Changelog
This site describes Keryn 0.5.2 and keryn-rules 0.8.0. The notes below are the release notes of the product, shortened to what changes for you: entries about the project’s own tests, tooling and internals are left out, and rules and edges an entry changed are linked. They are regenerated from the release notes at every build, so this page cannot disagree with them.
Keryn is the CLI, language server, apps and extensions; keryn-rules is the rule library.
Keryn (unreleased)
Section titled “Keryn (unreleased)”On the main branch, in no numbered release yet.
keryn aspm gate. : a check a pipeline can stop on, over the same comparisonkeryn aspm comparemakes.keryn aspm snapshot | compare | priority | report.keryn import runtimereads request-filter matches. :edgeevents from a Keryn Runtime agent become one INFO finding per rule, function, route and input, counting requests refused, requests that would be refused in monitor mode (or on a shadowed rule) and…
Changed:the eight edge kinds- the pattern rules: Go strings and
varspecs, and a faster join.
Changed:pack llm-output-taint-js-v1,seceng.llm01.taint.model-output-to-system-prompt-js - General injection rules for JavaScript/TypeScript and Python web frameworks. : new pack
general-injection-javascript-v1(a request value from Express, Koa, Fastify, NestJS or Next.js reaching SQL text, a shell command, a file path, an outbound URL, a redirect target or an HTML response) and…
Changed:pack general-injection-javascript-v1,pack general-injection-python-v1 keryn forge scan. : one command over compiled Java or.NET.- Maven and Gradle without a lockfile. :
keryn scanreads a resolved dependency list the build already produced, like a lockfile, in Keryn’s own matcher. keryn sca image <reference> --pull. : scans an image by registry reference with no daemon and no docker, skopeo or crane.- Organisation secret rules and
.gitleaksignore. : the secrets stage takes a gitleaks TOML config:--secrets-config <file>(scan, ci),.keryn.ymlsecrets.config(relative to that file, never outside its directory), or a at the scanned root when neither is set. - Secrets in git history. :
keryn scan --secrets-history [--secrets-since <ref|date>] [--secrets-max-commits N](andkeryn ci) also scans the added lines of every commit reachable from HEAD (git log -p -U0, the stream gitleaks’ owngitmode reads). - Design system, remaining charts and disclosures. : the organization’s “Open blocking cases over time” chart sits in a
ChartFramewith its table, in--chart-1; the decision desk’s three raw<details>areDisclosureand its title is on the type scale, so no raw<details>or… - Design system across the remaining views. : every show/hide in the app is now a
Disclosure(the 39 raw<details>left after the results views; aplainvariant for one inside a card or row,flushbodies and a controlledopen), the posture radar sits in a… - The remaining dashboard views render from the metric contract. : the Scan dashboard’s “What was checked” takes each capability’s unit, planned and completed counts, state and beside-the-denominator counts from
capabilityCoverage(a workflow row appears when a workflow analysis is supplied). - Named evidence milestones. : the organization view shows time from a blocking case’s first observation to its first cross-file evidence (proof 3) and to its reproduction (proof 5).
- The web app and VS Code open an entitlement-keyed rule vault. : signed in to Keryn Cloud, the web app gets an entitlement whose vault key is wrapped to a per-tab device key, verified and unwrapped with WebCrypto, and carried to the scan and Deep workers.
- Recent scans show branch, commit and labelled counts. : the Scan home’s Recent scans (local history, the last ten) now gives each row its branch, short commit and counts (errors, warnings, info; AI-BOM inventory not counted), and marks the result that is open.
- Trends and a diff between any two scans. : a repository’s page compares any two saved scans (introduced, newly detected, reopened, fixed only on a comparable rescan, gone-not-comparable, unchanged, with the reasons).
- Explicit baselines; “fixed” needs a comparable rescan.
- Reachability for JavaScript/TypeScript, Python and Java dependencies.
- Registry licences with
--online. - The local-first contract is tested and enforced in the browser.
- A project without a lockfile no longer exits 3. : an unresolved declared manifest (
package.json, or Gradle without a lockfile) is named in coverage (“Unresolved dependency manifest: …; declared range, not resolved”) but no longer makes the run… - Every release decision names its versioned policy; exceptions can be made to expire. : the decision carries
policy: {version, gate, files}, where the version (kp1:<12 hex>) is a digest of the gate settings and the exact.keryn/decisions.yml,scope.ymland policy file. - Agentic triage cannot be steered by the scanned repository. : repository text addressed to the AI reviewer marks its finding untrusted.
- The plan’s local scans per day are enforced. : release builds stop at the entitlement’s
scansPerDayfor the organization’s UTC day, counting every device. - Large lockfiles are read by default. : a lockfile past
--max-file-size(1 MB by default) or the 2 MiB secrets cap is read for the dependency stage, up to 64 MiB, through the host’s newreadLarge; the pattern and secrets passes keep their caps and never see it. - SCA severity: CVSS v4 and v2 scored, unscored advisories are unknown, not info.
- An offline scan says its dependency result checked nothing. : when the tree has lockfiles and no advisory source (no
--online), or matching failed, the verdict headline readsKeryn — Checks passed (Go); dependency advisories not checked (offline), the reasons and theAnalysis:line say… - Dependency SBOM: CycloneDX 1.6 and SPDX 2.3. :
keryn sbom [path]writes every package the tree’s lockfiles lock (no rules, offline, lockfiles up to 64 MiB), andkeryn scan --sbom <file>writes the same document from a full scan;--sbom-format cyclonedx|spdx… - Keryn scores: Trust and Risk.
- Dependency licences and a licence policy. : every
keryn scannow records each locked package in the result JSON underdependencies(offline too; only advisory matching needs--online): per lockfile, its purl, direct/transitive, scope, introducing path, lockfile line… - Scan dashboard. : the first screen of a result is six components: Release assessment, Needs your attention (blocking cases, cases to investigate, exceptions),…
- “Plan fixes” shows control candidates first; one Scenario estimate. : the trust map’s Plan panel lists control candidates without asking for any number.
- Organization view: action queues replace the risk score. : the portfolio no longer ranks repositories by
error×10 + warning×3 + info×0.5 + proven×15. - Profile sources scanned as part of their repository. :
keryn scan <dir> --repo-root <repo>shows the folder’s files under their repository path, so rulepaths:globs, code roles,--include/--exclude, the repository’s.keryn.yml(read from the repository root) and Deep see… - Deep time limit per profile source.
sources[n].deepTimeout(seconds, 1 to 86400) gives one source’s Deep pass longer than the 300 s default (keryn scan --deep-timeout <seconds>;KERYN_DEEP_TIMEOUT_MSstill works, the global default is unchanged). keryn sca scanhonours.keryn.ymlroles:.roles.suppresschooses which dependency findings are set aside (from--config, else<repo-root>/.keryn.yml, else./.keryn.yml;--no-configkeeps the defaults).keryn cilane: honest dependency stages, provenance, determinism, a stage pool,sca sbom. : a profile source with no manifest and no lockfile gets no dependency stage (the plan andkeryn application showsay “no dependency inputs”; a manifest without a lockfile still gets one).keryn artifact analyzelane: large executables, components from binaries, signature kind, unknown formats, extension manifests. an ELF, Mach-O or PE is parsed whole up to 256 MiB; past that, or past an explicit--max-artifact-size, it is read in 2 MiB windows (streamed hashes, string runs, the crate/runtime/signature readers over every byte), structure is not…- Application profile × packages, Deno and workspaces; reproducible ledger bytes.
- One merged findings ledger per run. :
keryn ciwritesledger.json(keryn.dev/ledger/v1,schemas/keryn-ledger.schema.json), one entry per distinct finding across source, secrets, dependency, web and binary stages. - Binary analysis looks inside release packages. :
keryn artifact analyze <package>unpacks.deb (ar + control/data tar, any of gzip/xz/zstd/bzip2),.rpm (lead + headers + cpio payload), AppImage type 2 (ELF runtime + squashfs payload) and plain… - the pattern rules’ taint facts reach Forge.
- Forge Java producer: library callbacks and DTO getters. an override of a library type’s method (jjwt
SigningKeyResolver) takes its parameters’ data category from the shipped source models, and a getter on an in-scan DTO returns what its receiver carries, sovulnerable-jwt-kid-callbackand… - Guard semantics on Forge. (workstream B3/B4): a sink category declares which guard classes clear it (
categories::clears_with; a sanitizer or guard of another class iswrong_contextand never clears, a value allowlist does not clear a deserialization sink), and… - Native binaries as AI-BOM evidence.
- Supply-chain signals beyond CVEs, offline.
keryn scan --engine forge.- Go is analysed by
--engine forge. - Java is analysed by
--engine forge. - C# is analysed by
--engine forge. - C and C++ are analysed by
--engine forge, file-local. - Secrets, dependencies, agent skills and binaries are part of every scan (no flag, no configuration). runs four engines over the files the scan reads, side by side with it (they do not wait for Deep), in the CLI, the web page, the desktop app and the editors; cached reports already hold their findings.
- Passive web security: Rust ports of the OWASP ZAP passive scan rules, inside the scans that already exist (no new form, button or command).
Keryn 0.5.2, 2026-11-08
Section titled “Keryn 0.5.2, 2026-11-08”Behaviour changes and migrations
Section titled “Behaviour changes and migrations”-
Decision records are not scanned.
.keryn/decisions.yml,scope.ymlandreleases.jsonlare skipped like Keryn’s own outputs, so adding answers no longer changes the findings they answer; coverage says how many records were not scanned. Holds name the trust-model chain edge instead offile:line -> rule. -
keryn.app runs the whole canonical library, the same default as
keryn scan, instead of the risk and AI-inventory packs (about 40 rules). Expect more findings on the same repository in Keryn Web. -
The Cloud trust crawl honours robots.txt: disallowed addresses are skipped and listed in the result’s limitations and
robotsSkipped. A site behind a bot wall or sign-in wall readsblocked (HTTP 403), with no score and nothing reported as missing. The score reads “N of 48 trust artifacts evidenced”. -
Licence: publishing benchmark or comparison results is allowed (former section 3(h) removed).
-
Decision Desk on the web, and decisions in the hand-off reports.
#/decide, opened from a “Release decision” button: Holds with the trust-model chain edge and proof level, accepted risks, coverage and scope. -
Decision records are never scanned; Holds name the chain edge.
-
Release builds ship keryn.dev’s docs pack.
-
Benchmarking is allowed. The Keryn Software Licence no longer forbids publishing benchmark or comparison results (former 3(h)) or using Keryn to evaluate or benchmark another product (3(f) now covers developing, training or improving one).
-
Release decision.
keryn decide(Go / Go with accepted risk / Hold from a scan and.keryn/decisions.yml),keryn decide sign(a person signs; hash-chained.keryn/releases.jsonl; refuses in CI),keryn decide migrate [--write] [--remove-legacy]… -
Web and CLI run the same rules. keryn.app scans with the whole canonical library (1,398 rules), the same default as
keryn scan; it ran only the risk and AI-inventory packs (about 40 rules) before.
Keryn 0.5.1, 2026-11-07
Section titled “Keryn 0.5.1, 2026-11-07”Behaviour changes and migrations
Section titled “Behaviour changes and migrations”-
Cloud trust scans read more of a site: robots.txt, sitemaps and security.txt, the site’s
trust.*hosts, linked trust-centre platforms, the seed’s own first redirect to one new registrable domain, and a seed with a path is crawled in scope. A rescan of the same site can find artifacts it missed before. -
Deeper trust scans. The crawl follows robots.txt, sitemaps and security.txt, the site’s own
trust.*hosts and linked trust-centre platforms, in priority order with caps; a seed with a path is crawled in scope; the seed’s own first redirect to one new… -
Examples that work out of the box.
Keryn 0.5.0, 2026-11-07
Section titled “Keryn 0.5.0, 2026-11-07”Behaviour changes and migrations
Section titled “Behaviour changes and migrations”-
Unchanged targets come from the cache. A rescan of the same commit or file tree with the same rules, engine and settings returns the stored report without scanning. Use
keryn scan --no-cacheto force a fresh scan, andkeryn cache clear|stats|pathto manage the store. -
What a scan reads.
.gitignoreis respected (git-tracked files are always listed) and Keryn’s own earlier outputs are never scanned. Use--no-ignoreto scan ignored files as before. -
keryn initwritesrules: [bundled:all], the defaultkeryn scanuses, instead of the AI-BOM pack only. Configs written by an earlierinitstill name the AI-BOM pack; replace it to gate on vulnerability rules. -
No readable rule files ship. Rule packs and the Deep content pack are read from the encrypted rule vault (
wasm/keryn-vault_bg.wasm). Tooling that read.ymlrules fromrulepacks/ordeep-content/in an installed package no longer finds them, andkeryn rules lint bundled:<id>no longer prints rule text. -
Engine tranche. Everything sealed. Every shipped wasm module is sealed after linking; readable runs fell from 4,425 to 101 (all reviewed library text) across 15 modules, and
ip-gatefails any unsealed Keryn literal or unreviewed readable run. -
Findings from the docs session (2026-11-07).
-
Trust canvas: trust boundaries, chains and chokepoints on an edge-first graph, computed in Rust/WebAssembly.
-
Agent surface in the AI-BOM: which HTTP routes reach an LLM and which tools that agent can call.
-
AI-BOM shows logos, not words. Each component row carries its product or vendor mark (Vertex AI, Bedrock, Gemini, OpenAI, Hugging Face…), the vendor’s mark beside the product’s, and Lucide glyphs for kind and category; a component without a known mark gets its…
-
Web app on large repositories. The scanning card counts each pass on its own (pattern rules, known fixes, Deep: distinct files, never above the total, only growing; it read “4,126 of 1,663” when every pass’s reads were added up).
-
Rule vault: no readable rules or detection content in any artifact.
Keryn 0.4.0, 2026-11-05
Section titled “Keryn 0.4.0, 2026-11-05”Behaviour changes and migrations
Section titled “Behaviour changes and migrations”-
First public release: there is no earlier public version to migrate from.
-
Single binary loads the vendored rule sets.
-
Deep by default, in every shell.
-
CI adapters for GitLab, Azure Pipelines, Jenkins and Bitbucket.
-
Web page: Cancel. one action at a time, a Cancel button in the status row aborts the running scan through the worker protocol and the status reads “cancelled”; the status line is a polite live region.
-
File-based AI-BOM components. model artefacts on disk (
.safetensors,.gguf,.onnx,.pt,.ckpt,.pkl, …) becomemachine-learning-modelcomponents, agent instruction files (.cursorrules,CLAUDE.md,AGENTS.md, Copilot / Windsurf / Cline / Kiro rules)… -
Browser popup shows the BOM. components, policy verdict and the BOM caveat as text, shared with the web page (
bomSummaryText), next to the four downloads. -
Container image build repaired.
Keryn 0.3.1, 2026-10-27
Section titled “Keryn 0.3.1, 2026-10-27”- Fixed: the 0.3.0 artifacts report their tool version as 0.2.0. The 0.3.0 bump missed and the man page, so everything that reads
KERYN_VERSION(keryn version, the result’sversions.tool, the SARIF driver, the CycloneDX tool entry) in the v0.3.0 CLI, web, Chrome, VS Code and Tauri…
Keryn 0.3.0, 2026-10-27
Section titled “Keryn 0.3.0, 2026-10-27”- Rule library keryn-rules 0.7.3. (from 0.7.0, three tags the same day): the
benchmark-gaps-v1pack (14 Python rules for the flywheel-benchmark positives the pattern rules had no rule for), allow-list / approval / confinement guards on shell and path sinks, sanitiser-aware HTML…
Changed:pack benchmark-gaps-v1,seceng.mcp.command-tool
Keryn 0.2.0, 2026-10-24
Section titled “Keryn 0.2.0, 2026-10-24”- Evidence and the bundle. Every finding carries
evidence{path,lineStart,lineEnd,snippet,contentSha256}. - Deep-scan findings join the report.
- Rule library keryn-rules 0.7.0.
Keryn 0.1.0, 2026-10-22
Section titled “Keryn 0.1.0, 2026-10-22”- One AI-BOM, three observation bases. Components record how they were observed (
keryn:basis: observed-in-source, observed-in-lockfile-only, mentioned-in-docs-or-examples-only, observed-in-browser, observed-on-network) with attributes and signals; one CycloneDX 1.6 document… - Policy as code in Rust. (
keryn-bom-policy, inside the vendoredkeryn-authoritymodule) in every shell, with waivers, SARIF output, selectors by kind, category, provider, name, basis and attributes, and readerreferencesto framework ids. - AI stack of this site.
- Trust scan of the current tab. (extension) and
keryn trust scan: the workbench Trust Scanner without its crawler. - Scanner-report import. (
keryn import, 97 formats), workflow analysis (keryn workflow analyze, 23 engines), binary triage (keryn artifact analyze), runtime tool-call authority (keryn authority). - Framework references in every output. OWASP LLM Top 10 2025, OWASP Agentic Top 10 2026, CWE, NIST AI RMF 1.0 and MITRE ATLAS ids from rule metadata reach SARIF taxonomies, taxa and rule tags,
--summary-by, and policyreferences; worded as references, never compliance… - Security posture. MV3 manifest with
activeTab+scriptingonly, strict CSP, no telemetry, no eval, SHA-pinned workflows, every vendored engine digest-manifested and checked against its source repository. - Rule pack keryn-rules 0.3.0 + taxonomy contract 1.1.0 (926 rules, 97 packs).
Keryn 0.0.1, 2026-10-11
Section titled “Keryn 0.0.1, 2026-10-11”No entries listed for this version.
Rules (keryn-rules)
Section titled “Rules (keryn-rules)”keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
keryn.agents2026.pydantic-ai-tool-argument-to-execand-to-sql-or-filedeclare file-local scope andrequires_imports: ["pydantic_ai"]. Their@$AG.toolsource also matched FastMCP servers and in-house tool registries, where they reported Pydantic AI findings; FastMCP servers now get the MCP rules above instead.
Changed:keryn.agents2026.pydantic-ai-tool-argument-to-exec
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
keryn.general.python.sql-fstringno longer fires on a statement that binds its values through driver placeholders (?,%s,%(name)s,$1,:name) and interpolates only bare identifiers…. (FROM {table}, a column,LIMIT {k}): that is the parameterised form a fix moves to.
Changed:keryn.general.python.sql-fstring
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
keryn.general.js.sql-injection-request-to-querytakes one more sink:$DB.raw(text,...)where$DBwas made byknex(...),Knex(...),x.knex(...)orrequire('knex')(...)earlier in scope,…. whatever the variable is called.
Changed:keryn.general.js.sql-injection-request-to-query
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
No entries listed for this version.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
No entries listed for this version.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
java-spring-ai-v2.yml: seven of eight rules now run only in files that import the framework (analysis-scope: file-local,requires_imports). The three handler rules and the prompt-logging rule need a Java AI SDK import (Spring AI, LangChain4j, OpenAI, Azure OpenAI, Anthropic, Google GenAI, Vertex AI, Bedrock Runtime); the four@Toolrules need Spring AI or LangChain4j, the…
Changed:pack java-spring-ai-v2- The Semantic Kernel pack stays ungated in this round: C#
global usingand project-level usings put a namespace in scope without a line in the file, so a file-local import gate could hide real…. findings.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- Added a Python SAST rule for a direct loopback-based administrative identity grant, with negative controls for authenticated access.
- Refreshed the six bundled OSV malicious-package catalogs, including the November Microsoft Agent Framework releases.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- New pack
general-injection-javascript-v1.yml: six taint rules from an HTTP request to a dangerous call, for JavaScript and TypeScript:keryn.general.js.sql-injection-request-to-query,….command-injection-request-to-shell,path-traversal-request-to-file,ssrf-request-to-outbound-url,open-redirect-request-to-location,xss-request-to-html-response.
Changed:pack general-injection-javascript-v1,keryn.general.js.sql-injection-request-to-query general-injection-python-v1.ymlgains the same six rules for Flask, Django (and REST framework) and FastAPI (keryn.general.python.*-request-to-*). The issue asked for them in this pack, so this is the one place where a rule was added to an existing pack file.
Changed:pack general-injection-python-v1- Both packs treat a helper whose name says it validates a file name, path or URL (
sanitizeFileName,get_safe_url,…) as a sanitizer, and a preceding prefix or..check as a guard. That is a naming heuristic: it lowers fires-on-fix and can hide a flow through a helper that does not do what its name says.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
No entries listed for this version.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- Inventory (P3): every code-form
keryn.inv.*rule ininventory-aibom-v1/v2reads source and dependency manifests only (paths.include; model/service rules also config files); bare package names…. became import/manifest forms. opt-in/: packs linted and tested like root packs but not part ofbundled:all(README, AUTHORING).- Rule bugs (P20):
seceng.gateway-localai-uploads-enabled-publicanchored to deployment config with a LocalAI gate (63 -> 0);seceng.llm05.rust-command-with-dynamic-argskips constant programs with…. constant arguments (63 -> 53, all remaining non-constant).
Changed:seceng.gateway-localai-uploads-enabled-public,seceng.llm05.rust-command-with-dynamic-arg,keryn.egress.py.secret-to-side-channel,seceng.mcp.command-tool - Inventory follow-up (monorepo pin check, lane qw/pin-fix): the line-anchored manifest forms (
^\s*"pkg"\s*:) missed one-line manifests ({"dependencies": {"openai": "^4.0.0",...}}), so the Keryn….
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
seceng.mcp.command-tool,seceng.mcp.filesystem-read-tool:requires_importsMCP server SDKs (@modelcontextprotocol/,mcp,fastmcp,rmcp,…). The bare wordsshell/exec/readFilematched every build script.
Changed:seceng.mcp.command-tool,seceng.mcp.filesystem-read-toolseceng.langgraph-human-interrupt-bypass:requires_importslanggraph,@langchain/langgraph.
Changed:seceng.langgraph-human-interrupt-bypasskeryn.general.js.trojan-source-bidi-control,keryn.general.python.bidi-control-character: embeddings, overrides and isolates only; the marks U+200E / U+200F / U+061C cannot reorder code and are…. normal in right-to-left locale copy.
Changed:keryn.general.js.trojan-source-bidi-control,keryn.general.python.bidi-control-characterkeryn.egress.{py,ts}.secret-to-side-channel: environment sources need a secret-like (or computed) name;os.environ.getis no longer its own request sink; a response’s parsed body and status are…. not the credential the request carried.seceng.cloudflare-worker-cors-wildcard-ai: not inside NodewriteHead, Puppeteerrequest.respond, Playwrightroute.fulfill, or a file importing Puppeteer / Playwright.
Changed:seceng.cloudflare-worker-cors-wildcard-aiseceng.github-actions-excessive-token-permissions-ai-job: the AI gate needs an AI word outside YAML comments.
Changed:seceng.github-actions-excessive-token-permissions-ai-job
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- Subprocess family (
general-injection-python-v1):subprocess-shell-dynamic-command(vuln),subprocess-shell-constant-command,subprocess-without-shell,shell-keyword-other-function,….process-partial-path(audit, per-file);shell=truthiness as Python reads it; findings on theshell=line.
Changed:pack general-injection-python-v1,pack general-hazards-python-v1 bind-all-interfacesalso reports"0.0.0.0"assigned, as a dict value and as a parameter default;hardcoded-password-stringalso reports password-named values set to"".model-supply-chain-v1:emit: per-fileonhf-from-pretrained-missing-revisionandhf-hub-download-missing-revision.
Changed:pack model-supply-chain-v1- Edge metadata re-derived for the new packs and
javascript-node-security-v1.
Changed:pack javascript-node-security-v1 vulnfix-gaps-v1:dill.Unpickler,jsonpickle.unpickler.decode/Unpickler,extractall(filter="fully_trusted"), DjangoSafeString/SafeText/SafeUnicode/SafeBytesand the fully qualified….mark_safe; a literal argument is recognised with any string prefix.
Changed:pack vulnfix-gaps-v1,pack model-supply-chain-v1prompt-redaction-ordering-v2: the logged variable must be named like a prompt or message (it reported any printed variable).
Changed:pack prompt-redaction-ordering-v2
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
No entries listed for this version.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- Audit rules report every place.
emit: per-fileon the Ruby, ERB, Haml and Slim audit rules. - *New
rails-dependency-versions-v1.yml(generic, Gemfile.lock / Gemfile / gems.rb / .gemspec /.ruby-version): Rails 7.1 or older, Ruby 3.0 or older, and json, loofah, rails-html-sanitizer,…. sprockets, actionpack-page_caching at versions with published advisories.
Changed:pack rails-dependency-versions-v1 - New
rails-hardening-ruby-v2.yml: RSA padding, catch-all routes, session cookie HttpOnly off or a literal session secret, hard-coded basic-auth credentials and….==password checks,show_detailed_exceptions?overrides andconsider_all_requests_local = truein production, auth/CSRF filters skipped with:except,:without_protection => true.
Changed:pack rails-hardening-ruby-v2 - SQL: Rails 2/3 finder option hashes,
update_all/delete_all/destroy_all, calculations; placeholder arrays, hash conditions, comparisons, permitted-parameter hashes and Arel predicates sanitise;…. built strings skip constants, table/key names, enum lookups, numeric literals, clock reads, the locale and quoting calls.
keryn-rules (unreleased)
Section titled “keryn-rules (unreleased)”On the main branch, in no numbered release yet.
- New
general-injection-php-v1.yml(10 rules) andsql-built-from-variableaudits for Java, Kotlin, C#, Swift and Go; insecure deserialization for Java, Kotlin and C#; wider request sources…. (Android, multipart, Web API, JAX-RS, Quarkus, Micronaut, Ktor); zip-slip entries identified by how they are obtained, guards on derived paths; ADOdb query API; PHP SSRF private-range guard.
Changed:pack general-injection-php-v1 - Framework gating:
metadata.requires_imports. (contract 1.2.0 -> 1.3.0, optional).
Changed:seceng.llm03.insecure-pickle-load,seceng.llm03.unsafe-model-deserialization - +33 rules for general bugs.
- Rules that still fired on fixed code.
Changed:seceng.wordpress-ai-output-echo-unsanitized,seceng.llm05.react-dangerously-set-inner-html,seceng.react.ai-output-to-link-href,keryn.vfgap.tarfile-extractall-unfiltered-py,seceng.llm01.python-concat-prompt,seceng.mcp.v3.tool-returns-env,6 more rules - Moved to inventory.
seceng.mcp.filesystem-read-toolandseceng.mcp.http-sse-serverdescribe a capability (their messages say “confirm”), not a defect.
Changed:seceng.mcp.filesystem-read-tool,seceng.mcp.http-sse-server
keryn-rules 0.8.0, 2026-11-07
Section titled “keryn-rules 0.8.0, 2026-11-07”metadata.edgeon every Keryn rule (1,329), metadata only: no pattern, message, severity, confidence or language changed. The value is one of the Edge Doctrine’s eight edge kinds,reaches,calls,reads,writes,sends,grants,depends_on,trusts, ornonewhen the finding is a property of one thing rather than a relationship.
Changed:edge reaches,edge calls,edge reads,edge writes,edge sends,edge grants,edge depends_on,edge trusts,the eight edge kinds,boundary user->prompt,boundary model->code,boundary secret->outside,boundary supply-chaininventory-aibom-v2.yml+8 inventory rules. (71 -> 79; INFO, capability, generic):keryn.inv.lib.gradio,.bentoml,.mlflow,.ultralytics,.spacy,.onnx,.dvc,.ray.
Changed:pack inventory-aibom-v2,keryn.inv.lib.gradio,keryn.inv.lib.bentoml,keryn.inv.lib.mlflow,keryn.inv.lib.ultralytics,keryn.inv.lib.spacy,keryn.inv.lib.onnx,2 more rulesinventory-aibom-v2.yml+16 inventory rules (79 -> 95): AI technologies our own security packs already target but the AI-BOM could not name. Agent frameworks:keryn.inv.lib.claude-agent-sdk,.google-adk,.strands-agents,.bedrock-agentcore,.microsoft-agent-framework,.agno.
Changed:pack inventory-aibom-v2,keryn.inv.lib.claude-agent-sdk,keryn.inv.lib.google-adk,keryn.inv.lib.strands-agents,keryn.inv.lib.bedrock-agentcore,keryn.inv.lib.microsoft-agent-framework,keryn.inv.lib.agno,10 more rules
keryn-rules 0.7.12, 2026-11-05
Section titled “keryn-rules 0.7.12, 2026-11-05”No entries listed for this version.
keryn-rules 0.7.11, 2026-11-05
Section titled “keryn-rules 0.7.11, 2026-11-05”No entries listed for this version.
keryn-rules 0.7.10, 2026-11-04
Section titled “keryn-rules 0.7.10, 2026-11-04”- Licence: the Keryn Software Licence.
keryn-rules 0.7.9, 2026-11-04
Section titled “keryn-rules 0.7.9, 2026-11-04”No entries listed for this version.
keryn-rules 0.7.8, 2026-11-03
Section titled “keryn-rules 0.7.8, 2026-11-03”No entries listed for this version.
keryn-rules 0.7.7, 2026-11-03
Section titled “keryn-rules 0.7.7, 2026-11-03”- Inventory precision and recall after the model-labelled pass.
Changed:pack inventory-aibom-v1,keryn.inv.service.endpoint,keryn.inv.lib.google-genai,keryn.inv.lib.azure-ai,keryn.inv.model.openai,keryn.inv.vectordb,keryn.inv.lib.huggingface,5 more rules
keryn-rules 0.7.6, 2026-10-27
Section titled “keryn-rules 0.7.6, 2026-10-27”- Sixteen regex-only rules on code languages now match the syntax tree.
Changed:seceng.bundle-hardcoded-ai-key,seceng.wordpress-openai-key-hardcoded,seceng.neon.database-url-hardcoded,seceng.llm03.embedding-http-transport-contextual,seceng.bundle-tool-schema-exposed,seceng.bundle-embedded-system-prompt,13 more rules
keryn-rules 0.7.5, 2026-10-26
Section titled “keryn-rules 0.7.5, 2026-10-26”- Twenty
pattern-regexrules are linear on any file size, with identical findings.
Changed:seceng.laravel-ai-output-direct-db-write,seceng.manifest-remote-code-ai-extension,seceng.gitlab-ci-user-input-to-shell-or-ai,seceng.github-actions-secrets-to-issue-comment-llm,seceng.github-ai-pr-bot-commits-without-approval,seceng.forensics-ai-error-swallowed,10 more rules
keryn-rules 0.7.4, 2026-10-26
Section titled “keryn-rules 0.7.4, 2026-10-26”- Four file-level
pattern-regexrules no longer rescan the file per line.
Changed:seceng.bundle-source-map-with-ai-secrets-risk,seceng.bundle-chrome-message-handler-privileged-api,seceng.github-actions-excessive-token-permissions-ai-job,seceng.supabase-edge-cors-wildcard-with-service-role
keryn-rules 0.7.3, 2026-10-25
Section titled “keryn-rules 0.7.3, 2026-10-25”keryn.gaps.retrieved-text-in-system-role. reports the system-message dict only (the 0.7.2 call-level sink was tainted by retrieved text in the user message of the same call and fired on the CTF benign twinrag_langchain_prompt.py).
Changed:keryn.gaps.retrieved-text-in-system-role
keryn-rules 0.7.2, 2026-10-25
Section titled “keryn-rules 0.7.2, 2026-10-25”keryn.gaps.text-written-to-tempfile-then-executed. (new,benchmark-gaps-v1): tool or model text written to atempfile.NamedTemporaryFilethatsubprocessthen runs withsys.executable; the “argv is a literal list” shape hides the code execution because the payload travels through…
Changed:pack benchmark-gaps-v1,keryn.gaps.text-written-to-tempfile-then-executedkeryn.gaps.retrieved-text-in-system-role. also reports thechat.completions.create(...)call that carries the system message (the labelled line is the call), excluding calls whose system content is a constant.
Changed:keryn.gaps.retrieved-text-in-system-role,seceng.asi01.system-prompt-from-retrieved-or-tool-content-pythonkeryn.gaps.prompt-or-messages-logged-raw. is silent when a redaction call (*redact*) precedes the log line in the same function.
Changed:keryn.gaps.prompt-or-messages-logged-rawseceng.llm07.python-system-role-tampering. no longer fires on a system prompt built from string literals only (asseceng.llm07.persona-tamperingin 0.7.1).
Changed:seceng.llm07.python-system-role-tampering,seceng.llm07.persona-tampering
keryn-rules 0.7.1, 2026-10-25
Section titled “keryn-rules 0.7.1, 2026-10-25”- New pack
benchmark-gaps-v1. Written from the defect shapes of the labelled cases and the frameworks’ own APIs: retrieved text placed in the system role (taint fromindex.query/ retrievers /list[Document]parameters to a{"role": "system",...}dict), an…
Changed:pack benchmark-gaps-v1 - Shell sinks learn allow-lists and approval gates.
seceng.llm04.python-model-output-to-system-shell,seceng.llm05.python-subprocess-shell-true, the sinks ofseceng.llm05.taint.model-output-to-os-command-python/-to-subprocess-python,vulnfix-ai-ai_taint_commandand the two…
Changed:seceng.llm04.python-model-output-to-system-shell,seceng.llm05.python-subprocess-shell-true,seceng.llm05.taint.model-output-to-os-command-python,vulnfix-ai-ai_taint_command - HTML sinks accept a sanitised variable.
seceng.llm04.model-output-xss-jsandseceng.llm05.inner-html-assignmentare silent when the assigned value was produced byDOMPurify.sanitize/sanitizeHtmlearlier in the function.
Changed:seceng.llm04.model-output-xss-js,seceng.llm05.inner-html-assignment seceng.llm07.persona-tampering. no longer fires on a system prompt built from string literals only.
Changed:seceng.llm07.persona-tamperingseceng.tauri-config-broad-fs-shell-permissions. runs only ontauri.conf.json,capabilities/andsrc-tauri/paths (it is a text regex and fired on Python files that mention “shell” and “all”).
Changed:seceng.tauri-config-broad-fs-shell-permissions- Left as measured, deliberately:
keryn.mcp.py.tool-returns-fetched-content-unwrappedandseceng.llm05.python-sql-string-formatfire on two negatives whose “safe” label concerns a different…. property (token logging, a literal caller); the rules’ findings stand.
Changed:keryn.mcp.py.tool-returns-fetched-content-unwrapped,seceng.llm05.python-sql-string-format
keryn-rules 0.7.0, 2026-10-24
Section titled “keryn-rules 0.7.0, 2026-10-24”- Dead annotations removed. 29 annotation lines referenced 21 rule ids that exist in no pack: seceng.llm01.multi-modal-vision-injection, seceng.llm05.dangerously-set-inner-html-model-output, seceng.llm05.insert-adjacent-html-model-output,…
vulnUntestedis tracked and may only fall.- Strict CTF gate (A3).
- C2: 38 deprecated rules retired.
- C1: lint errors, not warnings.
Changed:pack ci-security-gate-fail-open-v1 - the pattern rules->Workbench validation contract.
keryn-rules 0.6.8, 2026-10-23
Section titled “keryn-rules 0.6.8, 2026-10-23”No entries listed for this version.
keryn-rules 0.6.7, 2026-10-23
Section titled “keryn-rules 0.6.7, 2026-10-23”ci-security-gate-fail-open-v1(new, 1, generic): a security scanner step (CodeQL, Semgrep, Snyk, Trivy, gitleaks, Bandit, Checkov, tfsec,…) set to continue-on-error / allow_failure in a GitHub…. Actions / GitLab / Azure pipeline — a decorative security gate.
Changed:pack ci-security-gate-fail-open-v1- **
agent-security-config-v1: paths broadened to Terraform/HCL (*.tf, .hcl) and env files (.env, .env), so the fail-open and JWT-none config rules also scan IaC and environment surfaces.
Changed:pack agent-security-config-v1
keryn-rules 0.6.6, 2026-10-23
Section titled “keryn-rules 0.6.6, 2026-10-23”- Result: 22 variants, 0 misses — broad multi-vendor empirical confirmation (superseding the earlier single-target-per-class check) and a regression guard against future rule drift.
keryn-rules 0.6.5, 2026-10-23
Section titled “keryn-rules 0.6.5, 2026-10-23”client-trusted-privilege-v1(new, 1, py/ts/js/php): a privilege / role / identity / scope field read from client input (request body/args/params or a PHP superglobal), in both dotted and subscript…. form.
Changed:pack client-trusted-privilege-v1,seceng.toolauth.policy-bypass-flag-from-request
keryn-rules 0.6.4, 2026-10-23
Section titled “keryn-rules 0.6.4, 2026-10-23”state-transition-integrity-v1(+1 → 8): TypeScript action-without-replay-guard (MCP tool handler with no nonce/idempotency key in its arg object). Replay now Python and TypeScript.
Changed:pack state-transition-integrity-v1
keryn-rules 0.6.3, 2026-10-23
Section titled “keryn-rules 0.6.3, 2026-10-23”agent-trust-assignment-v1(new, 2, vuln, py/ts): a session / token / identity / privilege / trust field assigned from model output or a tool/agent result — confused deputy (ARKX E7 “The Octopus”). Taint: model/tool-output source to an identity-named assignment target.
Changed:pack agent-trust-assignment-v1,seceng.toolauth.model-controlled-user-or-tenant
keryn-rules 0.6.2, 2026-10-23
Section titled “keryn-rules 0.6.2, 2026-10-23”canonicalization-order-v1(+2 → 4): Go (url.QueryUnescape / filepath.Clean) and PHP (rawurldecode / realpath). Now Python, TypeScript, Go, PHP.
Changed:pack canonicalization-order-v1state-transition-integrity-v1(+1 → 7): Go copy-without-owner-reapply.
Changed:pack state-transition-integrity-v1- Vendors: Python secret-egress now also matches urllib3-style
$P.request($M, $URL,...); the TypeScript JWT rule also matches josejwtVerify. Engine-limited to Python/TypeScript (tree-sitter grammar rejects the block pattern): the Java variants, and canonicalization/state-transition in languages beyond those listed.
keryn-rules 0.6.1, 2026-10-23
Section titled “keryn-rules 0.6.1, 2026-10-23”crypto-claim-integrity-v1(+3 → 15): Kotlin (Android) JWT “none” / ECB and Swift (iOS, CommonCrypto) ECB. Now spans Python, TypeScript, Java, Go, PHP, C#, Kotlin, Swift.
Changed:pack crypto-claim-integrity-v1state-transition-integrity-v1(+1 → 6): Go non-atomic check-then-debit.
Changed:pack state-transition-integrity-v1agent-security-config-v1(new, 2, generic): the config dimension — a guardrail/gateway config that defaults to allow/open on error (fail-open), and a token config that enables the JWT “none”…. algorithm.
Changed:pack agent-security-config-v1
keryn-rules 0.6.0, 2026-10-23
Section titled “keryn-rules 0.6.0, 2026-10-23”agent-guard-fail-open-v1(4, vuln): a security/approval gate whose exception or error path runs the side-effecting action instead of denying (default-allow). try/except with the action in the except branch and approval functions that return allow on error (Python); try/catch gate→action (TypeScript); an authorization error discarded with_(Go).
Changed:pack agent-guard-fail-open-v1canonicalization-order-v1(2, vuln): a value validated/allow-listed, then decoded or normalized (unquote, decodeURIComponent, realpath/resolve), then used at a path/command/URL sink — so the check…. saw different bytes than the sink.
Changed:pack canonicalization-order-v1state-transition-integrity-v1(5): a copy/clone that re-applies no owner/ACL to the destination (audit); a non-atomic balance/quota check-then-debit (TOCTOU double-spend, vuln); a side-effecting…. agent/MCP/A2A action handler with no nonce/idempotency parameter (audit).
Changed:pack state-transition-integrity-v1crypto-claim-integrity-v1(12, vuln): a JWT whose signature is not verified (alg “none”, verify_signature/RequireSignedTokens false, UnsafeAllowNoneSignatureType, Algorithm.none) and claim/identity…. data encrypted with ECB (independent-block cut-and-paste).
Changed:pack crypto-claim-integrity-v1multichannel-secret-egress-v1(2, vuln): a secret (env/config/credential-named field) reaching an outbound request URL/body or a raised exception with no redactor on the path — redaction on the…. user-facing response is not enough.
Changed:pack multichannel-secret-egress-v1
keryn-rules 0.5.0, 2026-10-23
Section titled “keryn-rules 0.5.0, 2026-10-23”llm01-vendor-system-prompt-v1(6, taint): request data into Coherepreamble, Ollamasystem, and OpenAI Assistantsinstructions(Python and JS). Gemini, Mistral, Bedrock and Anthropic system sinks were already covered byllm01.taint.request-to-system-prompt-*.
Changed:pack llm01-vendor-system-prompt-v1model-loading-v1(14): Kerassafe_mode=False, NumPyallow_pickle, pandasread_pickle, cloudpickle/dill, skops trust-everything,torch.package, native extension loading,torch.hubwithout….weights_only/ withtrust_repo,torch.jit.load, unsafeyaml.loadon model configs,hf_hub_downloadresults unpickled, and a taint rule from request parameters into model loaders.
Changed:pack model-loading-v1vendor-sdk-safety-controls-v1(10): OpenAI calls without moderation and moderation results read withoutflagged, GPT Actionsx-openai-isConsequential: falseon mutating operations, Gemini…. safety thresholds set to BLOCK_NONE/OFF, Vertex AI resources without CMEK and notebook/workbench instances with public IPs (Terraform).
Changed:pack vendor-sdk-safety-controls-v1mcp-tool-name-integrity-v1(5, Python + TS/JS + generic): MCP tool names containing non-ASCII or invisible code points (homoglyphrеad_file, zero-width joiner), case/separator lookalikes of the…. reference servers’ tool names (readFile,Read_File,create-pull-requestnext toread_file/create_pull_request), and client configs (.mcp.json, Claude / Cursor / VS Code / Gemini settings, Codex config) installing a server…
Changed:pack mcp-tool-name-integrity-v1
keryn-rules 0.4.0, 2026-10-23
Section titled “keryn-rules 0.4.0, 2026-10-23”third-party/agentic-semgrep-rules/(36 rules, MIT, basitalisandhu/agentic-semgrep-rules v0.2.0 at8923f41): model output to exec/shell/SQL/URL/path/HTML sinks, user input in system prompts, MCP….references/SOURCES.json: every source now carriesbundleDecision, the written reason it is or is not shipped. Never: semgrep-rules (Semgrep Rules License), opengrep-rules (Commons Clause), Trail of Bits, Decurity (CC-BY-NC-SA), kondukto (no licence).
keryn-rules 0.3.2, 2026-10-22
Section titled “keryn-rules 0.3.2, 2026-10-22”browser-agent-automation-v1.yml(22 rules, JS/TS + Python + JSON): model output driving Playwright / Puppeteer / CDP navigation, page scripts (evaluate, addScriptTag, Runtime.evaluate) and…. setContent HTML; page text in the system prompt; launches with the Chromium sandbox off, the same-origin policy off or TLS validation off; agents on the owner’s real browser profile or attached over CDP; host functions exposed to pages;…
Changed:pack browser-agent-automation-v1rust-agent-desktop-v1.yml(13 rules, Rust + JSON): model output reaching process commands, SQL strings, filesystem paths, reqwest fetches, desktop open handlers and webview eval in Rust agents;…. reqwest clients with TLS verification off or without a timeout; Tauri v2 config with CSP null, dangerousRemoteDomainIpcAccess or withGlobalTauri; capabilities granting shell spawn/open, whole-filesystem or unrestricted HTTP reach, or…
Changed:pack rust-agent-desktop-v1agent-outbound-actions-v1.yml(6 rules, JS/TS + Python): model output as the recipient of an email, SMS or Slack message (Resend, Nodemailer, SendGrid, Twilio, Slack, smtplib), as the body of a….
Changed:pack agent-outbound-actions-v1fastapi-ai-service-v1.yml(6 rules, Python): CORS wildcard origin with credentials (FastAPI/Starlette and Flask-CORS); model routes without an authentication dependency or guarding decorator…. (FastAPI and Flask); model routes without a rate limit; provider exception text returned as the HTTP error detail; uploads read whole before any size check.
Changed:pack fastapi-ai-service-v1agent-client-configs-v1.yml(4 rules, JSON): Gemini CLI autoAccept, sandbox off, MCP servers marked trust; VS Code chat.tools.autoApprove. Key names written from memory of those clients’ docs (flagged in the pack header).
Changed:pack agent-client-configs-v1agent-client-configs-generic-v1.yml(3 rules, generic): Codex CLI approval_policy “never” and sandbox_mode “danger-full-access” (config.toml); Claude Code SKILL.md allowed-tools pre-approving Bash…. / Write / Edit unrestricted.
Changed:pack agent-client-configs-generic-v1
keryn-rules 0.3.1, 2026-10-22
Section titled “keryn-rules 0.3.1, 2026-10-22”- 17 rules listed
javascriptnext topython: Semgrep reuses the JavaScript parse of a pattern for every language of such a rule, so their Python alternatives (keyword arguments, dict literals)…. never matched in native Semgrep. openai-assistants-hardcoded-thread:"thread_..."is not a string wildcard in Semgrep; the rule binds the argument and checks^["']thread_.claude-code-hooks-v1.yml(6 rules, settings.json): hook script by relative path, unconditionalpermissionDecision: allowfrom an echo,evalin a hook command, Stop hook that blocks without…. checkingstop_hook_active,ANTHROPIC_BASE_URL/OPENAI_BASE_URLoverride inenv, DNS lookup with a variable in a hook.
Changed:pack claude-code-hooks-v1
keryn-rules 0.3.0, 2026-10-22
Section titled “keryn-rules 0.3.0, 2026-10-22”- 20 packs changed, patterns only (ids, messages, metadata, severities untouched).
Changed:seceng.workflow.expression-eval - : errors on top-level operators Semgrep ignores and on
eval/requirein mixed-language rules; comma-separatedruleid:lists count every id.
keryn-rules 0.2.0, 2026-10-22
Section titled “keryn-rules 0.2.0, 2026-10-22”- Known:
metavariable-regexis start-anchored in Semgrep CE; mid-string searches use a(?s).*prefix.
keryn-rules 0.1.0, 2026-10-22
Section titled “keryn-rules 0.1.0, 2026-10-22”- Severity policy: audit (absence) rules capped at WARNING; inventory rules INFO.