Project Glasswing claims OpenBSD findings as HN asks for CVEs and exploit proof

• Anthropic's Project Glasswing announcement says Claude Mythos Preview found an OpenBSD TCP SACK bug and many other issues, while the HN page card notes the company withheld details because more than 99% of findings were still unpatched.
• In the HN discussion summary, commenters pushed on the missing public evidence: no concrete CVE IDs for the OpenBSD claim, no exploit details for several cited findings, and no easy way to verify severity from the outside.
• The same HN discussion summary also challenged Anthropic's wording around "vulnerabilities" when exploitability was not always shown, a distinction that matters more than the press-release framing.
• Simon Willison's write-up surfaced the core tradeoff behind the restricted rollout: Anthropic is treating Mythos as capable enough to hold back, but the public record is still thin on the exact bugs behind that decision.

Anthropic's research post on Mythos Preview claims a crash-worthy OpenBSD TCP SACK issue, broad zero-day discovery across major operating systems and browsers, and a coordinated-disclosure backlog large enough that most findings stay private. The main HN thread immediately zoomed in on the missing CVEs, the gap between bug discovery and demonstrated exploitation, and whether the OpenBSD result was a security vulnerability in the strict sense or a less explosive reliability bug.

Anthropic's public claim is specific enough to be notable and still too thin to audit. In the Mythos Preview post, the company says the model found a vulnerability in OpenBSD's TCP SACK implementation that could crash any host responding over TCP, after roughly 1,000 runs.

According to the HN discussion summary, the follow-up question was simple: where are the CVEs, patches, or concrete identifiers for that OpenBSD bug and the other headline findings? Anthropic's answer, in the same research post, is that over 99% of discovered issues remain unpatched under coordinated disclosure, which explains the silence but does not resolve the verification gap.

The sharpest pushback was about terms, not just receipts. In the HN discussion summary, commenters argued that calling something a vulnerability without showing exploitability can blur together very different classes of bugs.

That critique lands against Anthropic's own framing in the official research post, which pairs the OpenBSD claim with broader statements that Mythos can identify and exploit zero-days across major operating systems and browsers. The public materials do not tie each named finding to a working exploit, so the HN thread ends up separating two claims that the launch copy often presents together: bug finding, and bug exploitation.

Glasswing is also a deployment story, not just a capability story. The official Glasswing announcement says Anthropic gave access to a select group that included AWS, Apple, Google, Microsoft, NVIDIA, and more than 40 other organizations, backed by $100 million in usage credits and $4 million in donations to open-source security groups.

Simon Willison's summary adds the part Anthropic clearly wanted readers to take seriously: Mythos was held back from general release altogether. That makes the absent bug-level evidence more conspicuous, because the public justification for restricting the model rests on findings that outside researchers still cannot inspect in detail.