Tracking Anthropic-discovered vulnerabilities
AI systems like Anthropic’s Glasswing/Mythos are reportedly generating tens of thousands of vulnerability findings in open source and proprietary software, but only a small fraction are being fixed or clearly exploited, raising doubts about impact, prioritization, and marketing hype. Commenters highlight high false-positive and low-severity rates for many reports, as well as the burden this places on already-stretched maintainers, while others say the tools have uncovered serious, non-public flaws inside large companies. The broader tension is whether frontier models are meaningfully improving security or mainly overwhelming human teams with noise, and how to balance early detection against limited capacity to verify and remediate issues.
Overall impact of Glasswing / Mythos findings
- Many see the public stats (≈38k findings, ≈7k public, ≈200 fixed) as showing AI is far better at finding issues than producing fixes or driving remediation capacity.
- Others argue this is a biased slice: the “ledger” is mostly OSS under coordinated disclosure, while the main Glasswing use is against proprietary code where fixes are reportedly triaged and applied internally.
- Several note that the metric “exploited in the wild” is a poor proxy here, since the point of early discovery is to preempt exploitation.
PHP2Shell and “catastrophic exploit” expectations
- Long debate over the PHP vulnerability: some call it a dramatic but manageable bug; others argue it was serious, underappreciated, and poorly covered by an early skeptical blog post.
- The “stealmybitcoin” experiment (vulnerable PHP server with a private key) was successfully exploited in under a day, apparently using an AI agent, cited as proof of exploitability.
- Still, there has been no visible “apocalypse” of PHP-driven mass breaches, with explanations ranging from rapid emergency patching to real-world exploit complexity.
Quality, false positives, and maintainer burden
- Multiple maintainers report high false-positive rates or “technically true but practically low-impact” issues, especially from the OSS scanner; numbers mentioned range from ~50% false positives to only ~30% truly important.
- Others report the opposite: batches where every finding was a real bug, with excellent test cases but patches needing human refinement.
- Even when accurate, AI reports often re-surface known bugs, minor hardening issues, or previously discussed problems, adding triage load without commensurate security gain.
Fixes, patches, and workflow
- Commenters note models can generate patches, and some Anthropic tooling already sends suggested fixes.
- Maintainers often find many of these patches low quality or risky, and say reviewing them can cost as much as writing a fix from scratch.
- There is broad concern that AI-generated “slop findings” overwhelm limited OSS maintainer time, pushing the need for better exploitability/risk scoring and prioritization.
Broader security landscape
- Several industry engineers say Mythos-class models have uncovered large numbers of serious issues in private codebases that will never get CVEs or blog posts.
- Some worry about LLMs chaining many “low-severity” flaws into real exploits; others emphasize that human error, not exotic 0-days, still drives many breaches.
- The article itself is widely viewed as aligned with the vendor’s business model (vuln prioritization), so some treat its negative framing as partially marketing-driven.