# AI-Accelerated Vulnerability Discovery in Core Runtimes

AI-assisted security research keeps surfacing critical bugs in long-trusted foundational code, so disclosures of this kind will recur more often and the time defenders have to patch will keep shrinking.

- Conviction: 43 / 100 (strengthening)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-11-18T00:00:00.000Z
- Last updated: 2026-11-22T06:36:09.478Z
- Canonical: https://polylog.news/ai/trends/ai-accelerated-vulnerability-discovery
- Publisher: Polylog
- Affected regions: Global

## Recent score history

- 2026-11-21: 40
- 2026-11-22: 43

## Recent evidence

- [confirms] Anthropic's Claude Fable 5.2 raises its FrontierCode 2.0 score to 49.1 percent and cuts API prices by 20 percent (2026-11-22): Claude Mythos 5.2, the restricted version of Fable 5.2 that runs the same underlying model with fewer safeguards, scores 89.6 percent on a benchmark for rediscovering known vulnerabilities. Near-saturation on bug rediscovery means AI-surfaced disclosures in trusted code should keep arriving faster.
- [confirms] Two HTTP/2 Flaws in NGINX Allow Unauthenticated Remote Code Execution on Static-File Servers (2026-11-19): Five research teams, including the AI security firm Metnir, independently found two HTTP/2 flaws in NGINX that allow unauthenticated remote code execution. One of the bugs has existed since HTTP/2 support was added in 2015. AI-assisted research is surfacing critical, long-hidden bugs in core infrastructure code, and several teams finding the same bugs at once suggests these finds are becoming routine.

2 more evidence entries, the full score history, the conviction-driver timeline, and affected assets are for subscribers: https://polylog.news/pricing
