Anthropic's scanner found 6 CVEs in Pion, some 4 years old

Pion's maintainers say Anthropic's OSS scanner found six security vulnerabilities in their DTLS library, some of them in code that had sat there for more than four years. They describe the results in a post on Ycombinator's Hacker News. The same scan found three bugs in a new DTLS API that OpenAI had already reviewed without flagging anything.

The maintainers added Pion DTLS to Anthropic's scanner in the week before October 17, through a pull request to the oss-scanner repository. On October 17 they woke up to a report listing 36 bugs. Six were security issues. The post says nothing more about the other 30. Four of the six vulnerabilities affect WebRTC, and the maintainers write that some had survived years of testing and security scans. They also note that all six predate ChatGPT. The published advisories:

• Critical: DTLS handshake signature confusion allows peer impersonation
• Critical: Client authentication bypass through premature session caching
• High: Client authentication bypass with shared session stores
• High: Unbounded handshake cache allows remote memory exhaustion
• High: Oversized RSA keys allow CPU exhaustion
• Moderate: Repeated certificate verification allows CPU exhaustion

Going by the advisory titles, the six split into two groups. Three are about identity, meaning who the peer is and whether a client actually authenticated, and both criticals are in this group. The other three are denial-of-service bugs, where a remote party can make the library use too much memory or CPU. Session caching comes up twice: one critical involves caching a session too early, and one high-severity bypass depends on shared session stores. In the Go package, client authentication is a policy the server sets through ClientAuthType, with modes from NoClientCert to RequireAndVerifyClientCert, according to the package documentation on Go's site.

The OpenAI review

The OpenAI comparison came out of a separate request. The maintainers had asked OpenAI to review a new DTLS API, because OpenAI uses Pion and had a use case for the API in SPED. OpenAI's review found no bugs. Anthropic's scanner found three in the same code: an authentication issue, a deadlock and a TLS KeyUpdate bug. The poster guesses that OpenAI's model may have skipped the KeyUpdate bug because it isn't a WebRTC issue. The API hasn't been released, and the fixes are in pull request 1236.

The people involved add some weight to the comparison. According to AI Engineer, Pion's creator and longtime maintainer Sean DuBois is a member of technical staff at OpenAI, and OpenAI's low-latency voice infrastructure runs on a Go service built on Pion. Still, this is one review set against one scan, and the post doesn't say which OpenAI model or setup did the review.

The maintainers also compare the scan with the AI-generated security reports that keep filling their inboxes, sent by people using Codex and Claude. They say most of those reports flag pointless issues, and none of them caught these bugs.

Who checks the findings

On October 11 we reported that Anthropic says it is "bottlenecked on our human capacity to validate these findings," and that OSS Scanner sends model-written reports straight to maintainers with no human triage. Pion's maintainers ended up doing that validation themselves. Neither Anthropic's nor OpenAI's models would help them verify or filter the 36 findings, so the team reviewed them by hand, one at a time, and it took the whole day.