Auth flaws let a free Claude account run models unbilled

A security researcher has published a write-up showing a chain of authorization flaws on Anthropic’s platform that let any signed-in claude.ai user, free tier included, run frontier Claude models without being billed. The main hole was quietly closed around September 1, but Anthropic had logged the report as platform abuse rather than a vulnerability, and the researcher left a related flaw — a paused paid subscription that still runs every model free — unreported. The post went out on Exploit on October 20, at the end of a 90-day disclosure window.
Underneath the bridge
The entry point is session bridging, a feature meant to let Claude Code hand a session over to Claude Desktop. A POST to /v1/code/sessions/{id}/bridge issued a signed worker token to any signed-in account regardless of plan, the one requirement being that the account had already sent at least one message. Anthropic’s own docs say the free Claude.ai plan does not include Claude Code access, so a free-tier account with chat-only capabilities was getting a credential built for a paid feature.
That token alone was not quite enough. Used against /v1/messages, a free account asking for a model like claude-fable-5 still drew a 429 and “This model is not available on your plan.” But the message batches API, /v1/messages/batches, applied no plan check at all. Any free account could reach every public model through it, and no account was billed for the work. The researcher files both gaps as CWE-863, incorrect authorization.
The catch was throughput. The batches API is built for bulk asynchronous jobs, up to 100,000 requests per batch processed at low priority, so each request took longer to come back, and the worker token had to be refreshed every eight hours.
Scoped as abuse, not a bug
Anthropic closed the report as Informational and treated the chain as platform abuse, a label normally reserved for things like using a trusted domain to push phishing or malware. When the researcher asked for the scoping to be reviewed, the reply came from a bot, claudesec-h1, acting on the company’s behalf. Around September 1, despite the out-of-scope ruling, a product change made the batches endpoint start rejecting the worker tokens.
The classification is the sticking point, not the fix. Anthropic does rate and pay out on Claude Code flaws: it scored a separate Claude Code Desktop chain at CVSS 8.7 (High) and patched it, and it paid this researcher for an unrelated Electron bug 22 days after the report. The authorization chain is the one it declined to count as a vulnerability.
The part still open
On August 17 the researcher found a third instance of the same gap. A paid subscription that has been paused — his own, after he stopped paying for Max 20x on a frozen card — still reaches /v1/messages and runs every model free, and without the low-priority slowdown that came with the batches route. He has not reported it formally, citing Anthropic’s position that these are not vulnerabilities it scopes.
Up next
AI industryDeveloper toolsanalysis
Cloudflare’s Rust port of TypeScript runs up to 8x faster
Coding agents on big TypeScript repos spend their loop waiting on the type checker. Cloudflare’s Rust port, tsrs, runs up to 8x faster than TypeScript 7.0 on the same worker count. Would you put an unofficial Rust port in your CI, or stick with the official compiler?
Related stories
- Cathedral’s AI agents broke into Anthropic in under 12 hours
- A file over 256KB let Claude Code Desktop leak your keys
- Claude Max subscribers are auditing Anthropic’s meter
- Anthropic will bill again for requests its safeguards block
- Security researchers: stop blocking defenders
- Linear and Ramp users can pay for Claude without an API key
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

