Skip to content

Anthropic

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

4 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →

Anthropic’s OSS scanner reported 36 bugs in Pion DTLS, the Go library that handles encryption in Pion’s WebRTC stack. Six of them are security vulnerabilities: two critical, three high and one moderate. Pion’s maintainers disclosed the findings in a post on Y Combinator’s Hacker News on October 16. They wrote that some of the holes had been “sitting in the codebase for over 4 years, surviving years of testing and security scans.”

Four of the six affect WebRTC. One of the two critical ones is a handshake signature confusion that lets an attacker impersonate a peer. The maintainers also wrote that their inboxes are “constantly spammed with AI-generated security reports from people using Codex and Claude, most of which flag pointless issues,” and none of those reports caught these bugs.

The project enrolled in the scanner on October 10. After that, Anthropic runs it. The scanner’s README says it builds the project in an isolated VM, emails what it finds to the project’s contact, and that the reports “are model-generated and are not reviewed by a human.” Anthropic has admitted it is ”bottlenecked on our human capacity to validate these findings.”

Pion’s maintainers had also asked OpenAI to review a new DTLS API, because OpenAI uses Pion and had a use case for the API in SPED. That review flagged nothing. Anthropic’s scanner found three bugs in the same code: an authentication issue, a deadlock and a TLS KeyUpdate bug. The maintainers guessed the model may have skipped the last one “because it’s not a WebRTC issue.”

The three fixes went into a single pull request, merged to main. The new API, DetachedConn, hasn’t shipped in a stable release, so the maintainers wrote: “detachedconn isn’t released yet, So we’ll not issue CVEs for these issues.” The code is already in the v4.0.0-rc.6 pre-release.

Sorting through the report took all day, because neither company’s models would help check the findings. JoTurk, who enrolled the project, said in a Hacker News comment that both Claude and Codex “refused to help verify, triage, or develop proof-of-concept reproductions.” So the team went through the bugs one by one by hand.

For users on v3, version 3.1.13 (released at 09:14 UTC on October 16) fixes four of the advisories. The two authentication bypasses were fixed only in 3.1.14, released at 16:39 UTC the same day, so 3.1.14 is the release that covers all six. It caps peer RSA keys at 8,192 bits and the handshake cache at 1,000 entries. For github.com/pion/dtls/v2, every advisory lists versions up to 2.2.12 as affected, and none lists a patched release.

Our analysis

What we checked

On October 18 we went through pion/dtls on GitHub: its six security advisories, its releases, the fix pull requests and git blame on the lines the fixes replaced. We also ran the new regression tests at the commits that added the fixes.

How we check stories

The maintainers’ post calls the six bugs CVEs, but as of October 18 none of them has a CVE identifier: they are published GitHub advisories, and GitHub’s global advisory database doesn’t list them yet. The severity counts and the WebRTC impact hold up. The age claim is stronger than the post made it, since by the maintainers’ own affected-version ranges all six bugs are more than four years old.

AdvisorySeverityCVSSAffected since
Signature confusion, peer impersonationcritical9.12.0.0 (2020-04-03)
Premature session caching auth bypasscritical9.12.1.0 (2022-01-11)
Shared session store auth bypasshigh7.42.1.0 (2022-01-11)
Unbounded handshake cachehighnone2.1.4 (2022-05-10)
Oversized RSA keyshighnone2.0.0 (2020-04-03)
Repeated certificate verificationmoderatenone1.2.2 (2019-03-19)
  • Both session-caching bypasses say they don’t affect WebRTC users. The impersonation advisory says checking the SDP certificate fingerprint doesn’t stop the attack.
  • All six ranges start before ChatGPT’s launch on November 30, 2022. Blame dates give only a lower bound on age.
  • A September 2 comment on the DetachedConn pull request, posted under the handle zshang-oai, reads “this is the new Codex writing style” and adds “The content has been reviewed by a human to the best of his knowledge.” By our reading, the OpenAI review was Codex-assisted with human sign-off.
  • The new tests pass at the fix commits but don’t compile against the code before them, so we couldn’t show a fail-then-pass run, and we measured no CPU or memory figures.
  • The tracking issue opened at 07:09 UTC on October 16. Thirty fix commits, two releases and the Hacker News post followed within 14 hours.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.