Skip to content

    Type a page, a command such as keryn scan, or a rule id.

    Changelog

    View as Markdown

    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.

    On the main branch, in no numbered release yet.

    • keryn aspm gate. : a check a pipeline can stop on, over the same comparison keryn aspm compare makes.
    • keryn aspm snapshot | compare | priority | report.
    • keryn import runtime reads request-filter matches. : edge events 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 var specs, 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 scan reads 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.yml secrets.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] (and keryn ci) also scans the added lines of every commit reachable from HEAD (git log -p -U0, the stream gitleaks’ own git mode reads).
    • Design system, remaining charts and disclosures. : the organization’s “Open blocking cases over time” chart sits in a ChartFrame with its table, in --chart-1; the decision desk’s three raw <details> are Disclosure and 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; a plain variant for one inside a card or row, flush bodies and a controlled open), 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.yml and 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 scansPerDay for 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 new readLarge; 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 reads Keryn — Checks passed (Go); dependency advisories not checked (offline), the reasons and the Analysis: 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), and keryn 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 scan now records each locked package in the result JSON under dependencies (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 rule paths: 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_MS still works, the global default is unchanged).
    • keryn sca scan honours .keryn.yml roles:. roles.suppress chooses which dependency findings are set aside (from --config, else <repo-root>/.keryn.yml, else ./.keryn.yml; --no-config keeps the defaults).
    • keryn ci lane: 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 and keryn application show say “no dependency inputs”; a manifest without a lockfile still gets one).
    • keryn artifact analyze lane: 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 ci writes ledger.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, so vulnerable-jwt-kid-callback and…
    • 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 is wrong_context and 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).
    • Decision records are not scanned. .keryn/decisions.yml, scope.yml and releases.jsonl are 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 of file: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 reads blocked (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.

    • 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.

    • 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-cache to force a fresh scan, and keryn cache clear|stats|path to manage the store.

    • What a scan reads. .gitignore is respected (git-tracked files are always listed) and Keryn’s own earlier outputs are never scanned. Use --no-ignore to scan ignored files as before.

    • keryn init writes rules: [bundled:all], the default keryn scan uses, instead of the AI-BOM pack only. Configs written by an earlier init still 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 .yml rules from rulepacks/ or deep-content/ in an installed package no longer finds them, and keryn 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-gate fails 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.

    • 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, …) become machine-learning-model components, 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.

    • 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’s versions.tool, the SARIF driver, the CycloneDX tool entry) in the v0.3.0 CLI, web, Chrome, VS Code and Tauri…
    • Rule library keryn-rules 0.7.3. (from 0.7.0, three tags the same day): the benchmark-gaps-v1 pack (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
    • 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.
    • 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 vendored keryn-authority module) in every shell, with waivers, SARIF output, selectors by kind, category, provider, name, basis and attributes, and reader references to 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 policy references; worded as references, never compliance…
    • Security posture. MV3 manifest with activeTab + scripting only, 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).

    No entries listed for this version.

    On the main branch, in no numbered release yet.

    • keryn.agents2026.pydantic-ai-tool-argument-to-exec and -to-sql-or-file declare file-local scope and requires_imports: ["pydantic_ai"]. Their @$AG.tool source 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

    On the main branch, in no numbered release yet.

    • keryn.general.python.sql-fstring no 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

    On the main branch, in no numbered release yet.

    • keryn.general.js.sql-injection-request-to-query takes one more sink: $DB.raw(text,...) where $DB was made by knex(...), Knex(...), x.knex(...) or require('knex')(...) earlier in scope,…. whatever the variable is called.
      Changed: keryn.general.js.sql-injection-request-to-query

    On the main branch, in no numbered release yet.

    No entries listed for this version.

    On the main branch, in no numbered release yet.

    No entries listed for this version.

    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 @Tool rules need Spring AI or LangChain4j, the…
      Changed: pack java-spring-ai-v2
    • The Semantic Kernel pack stays ungated in this round: C# global using and project-level usings put a namespace in scope without a line in the file, so a file-local import gate could hide real…. findings.

    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.

    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.yml gains 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.

    On the main branch, in no numbered release yet.

    No entries listed for this version.

    On the main branch, in no numbered release yet.

    • Inventory (P3): every code-form keryn.inv.* rule in inventory-aibom-v1/v2 reads 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 of bundled:all (README, AUTHORING).
    • Rule bugs (P20): seceng.gateway-localai-uploads-enabled-public anchored to deployment config with a LocalAI gate (63 -> 0); seceng.llm05.rust-command-with-dynamic-arg skips 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….

    On the main branch, in no numbered release yet.

    • seceng.mcp.command-tool, seceng.mcp.filesystem-read-tool: requires_imports MCP server SDKs (@modelcontextprotocol/, mcp, fastmcp, rmcp,…). The bare words shell / exec / readFile matched every build script.
      Changed: seceng.mcp.command-tool, seceng.mcp.filesystem-read-tool
    • seceng.langgraph-human-interrupt-bypass: requires_imports langgraph, @langchain/langgraph.
      Changed: seceng.langgraph-human-interrupt-bypass
    • keryn.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-character
    • keryn.egress.{py,ts}.secret-to-side-channel: environment sources need a secret-like (or computed) name; os.environ.get is 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 Node writeHead, Puppeteer request.respond, Playwright route.fulfill, or a file importing Puppeteer / Playwright.
      Changed: seceng.cloudflare-worker-cors-wildcard-ai
    • seceng.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

    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 the shell= line.
      Changed: pack general-injection-python-v1, pack general-hazards-python-v1
    • bind-all-interfaces also reports "0.0.0.0" assigned, as a dict value and as a parameter default; hardcoded-password-string also reports password-named values set to "".
    • model-supply-chain-v1: emit: per-file on hf-from-pretrained-missing-revision and hf-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"), Django SafeString/SafeText/SafeUnicode/SafeBytes and the fully qualified…. mark_safe; a literal argument is recognised with any string prefix.
      Changed: pack vulnfix-gaps-v1, pack model-supply-chain-v1
    • prompt-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

    On the main branch, in no numbered release yet.

    No entries listed for this version.

    On the main branch, in no numbered release yet.

    • Audit rules report every place. emit: per-file on 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 and consider_all_requests_local = true in 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.

    On the main branch, in no numbered release yet.

    No entries listed for this version.

    No entries listed for this version.

    • Licence: the Keryn Software Licence.

    No entries listed for this version.

    No entries listed for this version.

    • 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 twin rag_langchain_prompt.py).
      Changed: keryn.gaps.retrieved-text-in-system-role
    • 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,…
    • vulnUntested is 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.

    No entries listed for this version.

    • 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
    • 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.
    • 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
    • canonicalization-order-v1 (+2 → 4): Go (url.QueryUnescape / filepath.Clean) and PHP (rawurldecode / realpath). Now Python, TypeScript, Go, PHP.
      Changed: pack canonicalization-order-v1
    • state-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 jose jwtVerify. Engine-limited to Python/TypeScript (tree-sitter grammar rejects the block pattern): the Java variants, and canonicalization/state-transition in languages beyond those listed.
    • 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-v1
    • state-transition-integrity-v1 (+1 → 6): Go non-atomic check-then-debit.
      Changed: pack state-transition-integrity-v1
    • agent-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
    • 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-v1
    • canonicalization-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-v1
    • state-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-v1
    • crypto-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-v1
    • multichannel-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
    • llm01-vendor-system-prompt-v1 (6, taint): request data into Cohere preamble, Ollama system, and OpenAI Assistants instructions (Python and JS). Gemini, Mistral, Bedrock and Anthropic system sinks were already covered by llm01.taint.request-to-system-prompt-*.
      Changed: pack llm01-vendor-system-prompt-v1
    • model-loading-v1 (14): Keras safe_mode=False, NumPy allow_pickle, pandas read_pickle, cloudpickle/dill, skops trust-everything, torch.package, native extension loading, torch.hub without…. weights_only / with trust_repo, torch.jit.load, unsafe yaml.load on model configs, hf_hub_download results unpickled, and a taint rule from request parameters into model loaders.
      Changed: pack model-loading-v1
    • vendor-sdk-safety-controls-v1 (10): OpenAI calls without moderation and moderation results read without flagged, GPT Actions x-openai-isConsequential: false on 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-v1
    • mcp-tool-name-integrity-v1 (5, Python + TS/JS + generic): MCP tool names containing non-ASCII or invisible code points (homoglyph rеad_file, zero-width joiner), case/separator lookalikes of the…. reference servers’ tool names (readFile, Read_File, create-pull-request next to read_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
    • third-party/agentic-semgrep-rules/ (36 rules, MIT, basitalisandhu/agentic-semgrep-rules v0.2.0 at 8923f41): model output to exec/shell/SQL/URL/path/HTML sinks, user input in system prompts, MCP….
    • references/SOURCES.json: every source now carries bundleDecision, 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).
    • 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-v1
    • rust-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-v1
    • agent-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-v1
    • fastapi-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-v1
    • agent-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-v1
    • agent-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
    • 17 rules listed javascript next to python: 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, unconditional permissionDecision: allow from an echo, eval in a hook command, Stop hook that blocks without…. checking stop_hook_active, ANTHROPIC_BASE_URL/OPENAI_BASE_URL override in env, DNS lookup with a variable in a hook.
      Changed: pack claude-code-hooks-v1
    • 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/require in mixed-language rules; comma-separated ruleid: lists count every id.
    • Known: metavariable-regex is start-anchored in Semgrep CE; mid-string searches use a (?s).* prefix.
    • Severity policy: audit (absence) rules capped at WARNING; inventory rules INFO.