Source: raw/Black_October_Isn_t_Just_A_PHP_Bug._Attackers_Want_Your_AI_Keys..md — Creator: Nate B Jones, AI News & Strategy Daily · URL: https://www.youtube.com/watch?v=qDYe2eTl_D4 · Duration: not recorded · Platform: YouTube (fetched 2026-11-01). One same-week item (MLflow) is from raw/newsletter-theneurondaily-com-1fa0d4fd45.md (The Neuron Daily, October 28, 2026). Caption fixes: “Light LLM” is LiteLLM; “Clouder Codex” is “Claude or Codex”.
Jones’s argument is that AI API keys now behave like cash: whoever steals one runs models on your bill or resells the access. October 2026’s incidents (PHP, eMudhra, LiteLLM, Railway) showed how keys leak through services you depend on but never chose directly. He gives five steps for individuals, an incident walk-through for teams, and a test for picking the model that helps you investigate. He says a Substack post carries two audit prompts and board/CFO/engineering/vendor talk tracks; those were not part of this source.
Key Takeaways
- Keys are cash. A stolen Claude or Codex key lets someone “use AI on your dime,” run it themselves or sell it. When investigating an incident, check whether provider keys were reachable and whether they show unexpected usage or billing.
- You inherit risk from the whole chain. An app you connect to your files or calendar depends on a host and libraries you have never heard of; a weakness anywhere on that path can expose you.
- A clean repo is not proof of a clean install. LiteLLM’s October disclosure said the official source tag was clean while the distributed release contained credential-stealing malware (it had also been compromised in March). Ask “what code actually executed?”, not “what does the repo look like now?”
- Finished jobs still count. In his worked example the malicious release ran in a build job that has already exited and production is clean, but the job’s identity could read production provider keys, so those keys must be treated as exposed.
- AI tools can leak secrets too. LiteLLM reported seeing Claude Code put a user’s GitHub credential into a response that could end up in logs, and believes attackers took credentials from its deployment logs (still under investigation). He also relays a report of agents embedding user details such as an encoded email in URLs when sites ask agents to identify themselves; the user who asked Claude to look found about 10 such leaks over two weeks.
- Use the strongest model for defence. Offence and defence use the same actions (read unfamiliar code, write a test, run it, revise), so safeguards that block attackers also trip defenders. “This is not a place to be cheap.”
- Shrink what is reachable, not just patch faster. Anthropic reported almost 500 opportunities to cut exposed code or unsafe configuration combinations across 25 projects; in OpenSSH, Claude flagged signature-handling code reachable before login, which led to code removed in version 10.6.
What happened (as Jones reports it)
- PHP (fixes released October 21). Two critical flaws let an attacker run code ahead of a site’s own application under Apache’s mod_php. UK and European agencies told organisations to consider isolating or taking down systems they could not update. PHP’s maintainers had seen signs of one problem in 2024 and started and stalled the fix.
- eMudhra. Eight of the certificate authority’s private keys were trivially weak; one was the number four. Security researcher Scott Helme recovered them in under a second and says Claude did it from a single prompt. Because eMudhra was a trusted root, the mistake affected the whole chain of trust.
- Shields upstream. Mozilla pushed protection to standard Firefox configurations within a couple of hours of the report; Cloudflare and Vercel protected sites behind them.
- State actors. Microsoft documented a China state-aligned group using an AI agent framework (captioned “Artek”) against hundreds of organisations; most attempts failed, some reached privileged cloud access within hours.
- Railway. A small AI startup (captioned “Post”) took its service offline for a few hours on October 21 after warnings that its Anthropic and Stripe keys were compromised via the Railway breach. Jones calls that a good response: stop, regain control, document, then tell customers.
- Scale. Anthropic reported more than 29,000 candidate vulnerabilities from six months of scanning important projects, about 6,000 manually triaged so far. Bank of America took a charge of “a couple of hundred million dollars” last quarter to hire defenders.
Try It
If you use AI apps (his five steps):
- Update browser, phone, operating system and apps, and finish the restart. Ask Claude or Codex to check for updates across affected services when a breach is announced.
- Open connected-app settings for your email and file accounts (Google lets you review and remove third-party access). Remove unused connections; check what the rest can read or change. You can paste the app names and permission descriptions into your AI and ask it to flag risky ones.
- Remove unused browser extensions and, where Chrome supports it, restrict the rest to specific sites or on-click access.
- Use a password manager with unique passwords and multi-factor authentication everywhere; agree a voice passphrase with family against voice deepfakes.
- When a breach notice arrives, read what was exposed and act the same day: change passwords, revoke app connections, and if you are a developer rotate API secrets.
If you build with AI:
- Rotate every key a compromised build or host could reach, including keys only touched by jobs that have finished.
- Check what actually ran (deployment and build history, installed package hashes), not just the source repository.
- Pick your investigation model with a rehearsal. Recreate a past incident whose answer you know, give candidate models the same evidence and query tools without the answer, and see which ones get past the clean repo to the build record and connect the build identity to the credentials. Compare time and cost to the right answer.
- Audit agent output for leaked secrets and personal data, including logs and URLs your agents generate.
- Be transparent with customers: what you know, what is uncertain, whether access has stopped, and what they should do.
- Same week (The Neuron): MLflow patched a critical remote-code-execution bug; upgrade to 3.17.0.
Open Questions
- “Artek” (the agent framework in Microsoft’s report) and “Post” (the startup) are auto-caption spellings; neither was confirmed.
- Jones says “Blocktober” in the audio while the video title says “Black October”; this article uses the title.
- None of the incident details were checked against the PHP, LiteLLM, Railway or Microsoft advisories.
- The Substack prompts and talk tracks were not reviewed.
Related
- The Find-Fix Asymmetry — the PHP2Shell story and why fixing lags finding
- AI Search Weekly News, October 2026 — the same week’s incident list (Railway, LiteLLM, OpenCode, GitLab, eMudhra)
- Open-Weight Models and the Patch Gap — why a published fix is now a public exploit
- AI-Enabled Cyber Threats — attacker use of AI agents mapped to ATT&CK
- Claude Security Plugin — running Claude as the defender on your own code
- Redacting Files for Frontier Models (Nate B Jones) — keeping sensitive data out of model context