THE PERMISSION [LAYER]

Issue #4 · Tuesday, November 10, 2026

A wave of AI platform breaches is after model keys

A catch-up issue covering Oct 24 to Nov 10. Plus GitBook, MCP fixes and bank rules.

Get the next issue in your inbox

The Permission Layer is a free briefing on AI agent governance, sent at least once a week.

Free, at least once a week. You confirm by email.

This is a catch-up issue covering Oct 24 through Nov 10. The biggest story of the gap: attackers broke into AI developer platforms to steal the model keys customers store there.

THE WEEK IN CONTROL

GitBook says attackers “using AI agents” reached its production systems A flaw in GitBook’s SSO login let an attacker act as GitBook staff on Nov 2, reach internal repositories, then production secrets, with access lasting into Nov 3. GitBook calls it “a sophisticated, highly targeted attack using AI agents.” API tokens and integration secrets were exposed. Source Why it matters: Issue new GitBook API tokens, rotate integration secrets and re-authenticate its MCP server.

A contractor’s personal Claude Code agent reportedly ran on a Pentagon network WIRED reports that an Accenture employee connected a Defense Department system to a personal Claude Code account, and that the agent ran for two months before it was found. Accenture says the allegations were “thoroughly and independently investigated and found to be without merit.” Source Source Why it matters: Check whether your monitoring would catch an employee’s personal agent, not just the ones you issued.

A malicious RubyGem wipes the home directory when its stolen token is revoked Socket found that strutil-core 1.0.1, published Nov 1, is built to steal Claude Code and Codex credentials and plant an MCP server and agent instruction files. It then watches the stolen GitHub token: “Revoking the token is exactly what a responder would do first, and it is what triggers the deletion.” Source Why it matters: Isolate an infected developer machine before you revoke anything.

MCP SDK token fixes stay off until you configure them The official Python and TypeScript MCP SDKs both accepted tokens meant for other services: “a valid token for another service was also accepted by your MCP server.” Both advisories say “Upgrading alone changes nothing.” AWS’s API MCP Server had the same gap, fixed in 1.6.0. Source Source Source Why it matters: Check the configuration, not the version: validate_token_resource in Python, expectedResource in TypeScript, AUTH_AUDIENCE or AUTH_CLIENT_ID on AWS.

Langflow 1.13.0 is the only fix for two unauthenticated critical flaws IBM’s bulletin covers Langflow 1.0.0 to 1.12.8, including “unauthenticated remote code execution exploitable through model context protocol tool enumeration.” Two flaws score 9.8. Neither was on CISA’s exploited list as of Oct 30, but a public Metasploit module already lists a server’s MCP tools without credentials. Source Source Why it matters: Treat any reachable instance below 1.13.0 as exposed.

US bank examiners get a framework that covers AI agents The OCC’s Nov 4 examiner guide on AI, which is not binding on banks, ranks risk by what a system can do: “a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance.” Source Source Why it matters: If you sell agents to banks, expect to be asked how each one’s identity, credentials and logs are governed.

Three big Canadian banks block personal agents, and Wells Fargo wants proof of revocation RBC said on Oct 27 it “has decided to block third-party personal AI agents” from online banking; BMO and TD block them too. Wells Fargo’s terms, effective Nov 24, say of revoking an agent: “You must retain evidence of such revocation and must provide such evidence to us, at our request.” Source Source Source Why it matters: Products that need an agent inside a customer’s bank account need another route.

ONE DEEP THING

Attackers are breaking into AI platforms to spend your model budget

What happened Wiz’s running list names 16 AI developer platforms that have disclosed incidents since early October, among them Fireworks AI, Railway, Northflank, Lightning AI and Trigger.dev. Source Where details are public, the prize is the same: model provider keys that customers stored on the platform.

Railway’s full incident report, published Nov 10, says the attacker was “specifically targeting AI provider credentials” and that “Unauthorized spend on the stolen keys was estimated to reach ~$740,000.” Source Northflank now tells customers to assume every environment variable was decrypted: “Users have reported unauthorised use of AI API keys supplied to the Northflank platform.” Source Fireworks AI has widened its own disclosure to include billing data and, for a limited set of accounts, secrets, datasets and incomplete models. Source

Where the keys go Google’s threat team describes the market behind it. Stolen access is resold through what it calls LLM Mixers, which “function as illicit routing gateways that proxy, aggregate, and resell unauthorized LLM access to underground subscribers.” In one case a single reseller “passed nearly one million dollars in token costs to the victim organization.” Source

Why this is a governance problem A model key stored on a build or hosting platform is a spending credential held by a third party. Many carry no cap, no network restriction and no owner, so the first alarm is the invoice. In two cases Wiz tracked, stolen keys were first tested within 30 minutes of the breach.

What to check this week List every platform that holds one of your model provider keys, including the ones your vendors use. Rotate any key stored on a platform Wiz names. Set a hard spend cap at the provider on each key. Restrict where keys work: Anthropic’s new IP allowlist, in beta since Nov 5, “limits where a leaked API key can be used.” Source Alert on bursts of model calls, not just on monthly totals.

The takeaway: A model key is money. Give every key an owner, a cap and a network boundary before it goes onto someone else’s platform.

LANDSCAPE WATCH

  • CrowdStrike agreed to acquire Fiddler AI, an agent observability vendor, and Dreadnode, an AI security testing firm. Neither deal has closed. Source Source
  • Vanta signed an agreement to acquire Pillar Security, a runtime security platform for AI. Source

→ Full landscape map

NEW TOOLS

  • Okta for AI Agents, Okta (generally available): an agent registry with short-lived scoped access and a switch to deactivate an agent. Source
  • Execution Containers, Microsoft (generally available): limits on an agent’s files, network and processes, enforced by Windows from outside the agent. Source
  • AI Kill Switch, CrowdStrike (available at its AI Gateway; endpoint enforcement announced): stops a misbehaving agent and cuts its network access. Source
  • Inference hooks, Anthropic (generally available): send Claude Enterprise prompts and pending tool calls to your own security tools. Source

→ All launches from the last 90 days

SPEND SIGNAL

Two list prices moved in opposite directions. Anthropic’s Claude Fable 5.2 and Mythos 5.2, released Nov 9, “cost $8 USD per MTok of input and $40 USD per MTok of output,” below the 5.1 models. Source Google’s Gemini 4 Argon doubles on Dec 1: input goes from $2.00 to “$4.00 starting December 1, 2026” per million tokens, and output from $10.00 to $20.00. Source The tactic: rerun any agent budget built on Argon before December, and move caps with it.

README

Forward this to whoever holds your model provider keys, and reply with what your key inventory turned up.

The Permission Layer

Get the next issue by email

That was Issue #4. New issues go out at least once a week, with special editions when something big happens. Five-minute read, free.

You confirm by email. Unsubscribe in one click.