In separate reports, SecurityWeek describes a stolen access key, StepSecurity describes npm releases that steal secrets, and The Hacker News describes AI agents hunting for AI provider keys [3][5][33].
The point
SecurityWeek reports that the Cursor supply chain attack started at Fireworks AI, where the attacker stole an access key 3. StepSecurity reports that malicious WPPConnect npm releases steal secrets and WhatsApp sessions 5. The Hacker News reports that attackers' AI agents have been hunting for weak spots in internet-exposed AI apps such as LiteLLM, One API and Dify to steal the AI provider keys stored inside 33. This essay follows the key, secret or credential in each report and looks at what each source says it reached 3, 5, 33.
The evidence
The Fireworks key. SecurityWeek reports that the attacker stole an access key from Fireworks 3. According to SecurityWeek, the attacker used that key to exfiltrate credentials 3. SecurityWeek says those credentials enabled the attacker to push malicious Claude Code extension copies 3. SecurityWeek's excerpt does not say how the key was stolen 3.
The WPPConnect releases. StepSecurity reports malicious WPPConnect npm releases 2.4.1 and 2.10.82 5. According to StepSecurity, those releases steal secrets and WhatsApp sessions via Telegram 5. StepSecurity's headline says the releases steal secrets and WhatsApp sessions "through the project's own pipeline" 5. StepSecurity's write-up covers the affected versions, how they shipped and what to rotate 5.
The AI provider keys. The Hacker News reports that Amazon uncovered an AI agent campaign targeting LiteLLM, Dify and other apps for API keys 33. According to The Hacker News, the attackers' AI agents have been hunting for weak spots in internet-exposed AI apps to steal the AI provider keys stored inside 33. The Hacker News says this account comes from the agents' own records, which Amazon's threat intelligence team captured 33.
Where coding agents hold secrets. The authors of the arXiv paper "Agent Authenticator" write that AI coding agents now push code, publish packages, call model APIs and operate cloud resources on behalf of developers 19. They write that these agents authenticate with the same bearer secrets humans use: personal access tokens, API keys and package-registry tokens 19. The paper is titled "Agent Authenticator: Passkey-Grade, Hardware-Bound Authentication for AI Agents" 19. Separately, Dark Reading reports that a command in Claude Code inadvertently saves files with credentials and other sensitive data to public Git repositories 18.
What it means
We think none of these reports should be tied to the others, because none of the excerpts links them. We think each one is still worth reading for what the credential could reach. In our view, the Fireworks case shows how far a single stolen access key can travel. SecurityWeek describes a path from a stolen key, to exfiltrated credentials, to malicious Claude Code extension copies 3.
StepSecurity's headline uses the phrase "through the project's own pipeline" 5. We think that phrase is a reason not to treat a familiar publish path as proof of a clean release. In our view, consumers need release provenance they can check, and a matching package name and publisher are not enough.
The Hacker News reports that attackers' agents have been hunting for weak spots in internet-exposed AI apps to steal the AI provider keys stored inside 33. In our view, an AI app that holds provider keys is a credential store. We think it deserves the controls you would give a secrets vault, not the controls you would give a convenience proxy.
We think the "Agent Authenticator" paper describes why coding agents belong in this picture. Its authors write that agents push code, publish packages and call model APIs with the same bearer secrets humans use 19. In our view, whoever or whatever can read those secrets inherits those powers.
In our view, the trust boundary worth watching is any token that can publish code, publish a package or call a model API. Dark Reading reports that a Claude Code command inadvertently saves files with credentials to public Git repositories 18.
What to do next
- **Inventory the secrets your tools hold.** List every token held by coding agents, editor extensions, CI pipelines and AI gateways. For each, record an owner, its scope and the steps to revoke it.
- **Split publish rights from everything else.** Give package-registry and extension-marketplace tokens their own narrow scope. Keep them out of developer laptops and agent sessions where you can.
- **Act on WPPConnect now if you use it.** Search lockfiles and build caches for releases 2.4.1 and 2.10.82, remove them, and follow StepSecurity's guidance on what to rotate 5.
- **Treat coding-agent extensions as dependencies.** Pin versions, allowlist publishers, and review updates before they reach developer machines. Check that each extension comes from the publisher you expect before you install it.
- **Put AI gateways behind authentication.** Do not expose LiteLLM, One API, Dify or similar apps directly to the internet. Give each app its own provider key, set spend limits, and alert on unusual use.
- **Stop secrets before they reach a public repo.** Turn on push protection and pre-commit secret scanning for every repository a coding agent can write to. Check what your agent commands save and where.
- **Shorten credential lifetimes.** Prefer short-lived, scoped tokens over long-lived personal access tokens for agents. Evaluate hardware-bound options such as the authentication named in the title of the "Agent Authenticator" paper 19.
- **Rehearse revocation.** Pick one agent token and one registry token, then time how long it takes to revoke and replace them. We think that number tells you more about your exposure than any scanner report.
AI-drafted. Not reviewed by a human editor: it passed automated checks only. Published by Keryn. Corrections: keryn@keryn.dev
How this was checked
This essay was drafted by claude-opus-5-5 from the titles and excerpts of the sources listed; the full articles were not read. Before publication it passed four automated checks. (1) The Keryn voice scorer: grade 10.3, 886 words, no blocking finding. (2) A link check: every citation resolves to a listed source, and every number in a paragraph appears in the title or excerpt of a source that paragraph cites. (3) Claim checks: a count stays with the group its source gives it to, and a condition, a status (reported, alleged, under investigation, not verified) or an allegation in a source is kept. (4) A second-model review by claude-sonnet-5-5 of each sentence against those titles and excerpts: pass. These checks catch invented figures and unsupported sentences. They do not verify the sources themselves, and they cannot catch a fact the excerpts leave out. No human read this text before publication. Corrections are made on this page and dated here; write to keryn@keryn.dev.
Sources
- Optimizely supply chain attack infects major brands' sites with malware BleepingComputer, 17 November 2026
- Hackers compromise 4,800 GitHub accounts to steal cloud credentials BleepingComputer, 12 November 2026
- Cursor Supply Chain Attack Started at Fireworks AI SecurityWeek, 16 November 2026
- CVE-2026-118383: Use-after-free vulnerability in MongoDB Datadog, 11 November 2026
- WPPConnect Compromised: Malicious Releases Steal Secrets and WhatsApp Sessions Through the Project's Own Pipeline StepSecurity, 16 November 2026
- California sues OpenAI over AI agents escaping sandboxes, hacking firms BleepingComputer, 16 November 2026
- Preserve the Vulnerability, Break the Alert: Testing the Robustness of CodeQL Taint Tracking with Natural Rewrites arXiv, 12 November 2026
- How AI SAST Traced Data Flows to Uncover Six OpenClaw Vulnerabilities Endor Labs, 11 November 2026
- Endor Labs’ AI SAST Finds Zero Day Memory-Amplification DoS in Anthropic’s buffa library Endor Labs, 11 November 2026
- AI SAST vs traditional SAST: what changes when static analysis can reason Endor Labs, 12 November 2026
- BO Team Exploits Log4Shell in VMware vCenter, Compromises Russian Firm's Code-Signing Server The Hacker News, 16 November 2026
- Trained to betray: How post-training supply chain attacks hijack coding agents Microsoft, 12 November 2026
- 56 AWS API calls in 141 milliseconds: Inside an AI-run intrusion Elastic, 13 November 2026
- ARTEX and beyond: AI-enabled penetration testing platforms Cisco, 13 November 2026
- Adapting to AI penetration testing orchestrators: The Modern Security Model Cloudflare: Security, 13 November 2026
- OpenAI’s Daybreak brings AI agents to security testing and vulnerability fixes Help Net Security, 13 November 2026
- Compositional Exploits in Autonomous Coding Agents: Attack Mechanics and Ephemeral Defense arXiv, 15 November 2026
- AI Coding Agents Leak Thousands of Secrets to GitHub Dark Reading, 10 November 2026
- Agent Authenticator: Passkey-Grade, Hardware-Bound Authentication for AI Agents arXiv, 14 November 2026
- PrivacyPR: A Dataset of Real-World Privacy Fixes for Benchmarking Privacy-Aware Code Review arXiv, 13 November 2026
- Git security vulnerabilities announced GitHub, 10 November 2026
- N-able Patches 17 N-central Flaws as Attackers Exploit Earlier Bugs to Deploy Crypto Miners The Hacker News, 17 November 2026
- Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches The Hacker News, 16 November 2026
- AI uncovers flaws in discontinued devices with no vendor left to patch them Help Net Security, 16 November 2026
- Agent-ZTA: A chain-Aware Zero Trust Model for Bidirectional Security-State Propagation and Runtime Containment in Multi-Agent Delegation arXiv, 13 November 2026
- Isolation Is Not Confidentiality: Security Boundaries of Geometric Access Control in RAG arXiv, 10 November 2026
- Advancing a Global Framework for AI Safety and Governance for the Well-being of Humanity arXiv, 10 November 2026 · Background, first released 2025
- Comparing WeChat and Telegram MiniApp Security Architectures: Toward Cross-Platform Analysis arXiv, 10 November 2026
- 'Sleeper Agent' Tactic Hides Payloads in Open-Weight AI Models Dark Reading, 16 November 2026
- HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection arXiv, 15 November 2026
- China-Nexus Actors Breach 10 Organizations in AI-Driven Attacks Dark Reading, 16 November 2026
- Inside an AI Attack Against a US Retailer Dark Reading, 12 November 2026
- Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys The Hacker News, 16 November 2026