← Azimuth Daily
Capturing data…
Capturing data…
SUN, 25 OCT · 77 ITEMS
"Anthropic shows LLMs making formal software verification feasible at larger scale" — Hacker News (best) · AI & Technology
Anthropic has published research showing that large language models like Claude can make formal verification of software—mathematically proving that code is free of certain classes of bugs—practical at a much larger scale than previously possible. They applied the approach to 123 open-source libraries and uncovered hundreds of hidden vulnerabilities, suggesting that could dramatically reduce the cost and manual effort required for formal verification.
The claim is drawn from Anthropic's own primary research publication and blog post, linked directly from the Hacker News discussion. As a self-published result by the company, it is authoritative about what they did and found, but independent verification of the specific methodology and bug counts would require examining the full paper and reproducing the experiments.
This is genuinely new material published on October 22, 2026, not a resurfacing of older work.
We covered this development on October 24, 2026, noting Anthropic's formal verification results alongside reports of an IPO preparation.