Anthropic shares its recipe for using LLMs to find vulnerabilities

Anthropic published a detailed guide on running LLM-based vulnerability discovery at scale, written by Eugene Yan, Henna Dattani, and a group of engineers from the security team. The post breaks the workflow into six steps: threat modeling, sandboxing, discovery, verification, triage, and patching. Each step is described with the failure modes the team hit in production rather than a theoretical pipeline.

A few numbers stand out. Anthropic's own scanning has surfaced 1,596 vulnerabilities so far, of which only 97 have been patched. One partner team reports a 90 percent true positive rate on exploitable findings, but only after they wrote thorough threat models that scoped the search. Adding adversarial verification in a fresh container roughly halved their false positives, and requiring proof of concept execution pushed them close to zero. The high-level lesson is that discovery has become the easy part: horizontally scaling more agents saw limiting returns, while verification and triage are where the real bottleneck sits.

The prompting advice is also useful. Anthropic argues against prescriptive checklists and recommends short, goal-oriented prompts that hand context to the model rather than telling it how to look. They also warn that running verifiers in the same context as discovery agents creates confirmation bias, which is why they spin up independent containers without shared history.

Why it matters

If you are setting up an LLM security pipeline, budget for the work after the scan. The first run will return more findings than you can act on, and you will spend more engineering on triage and patches than on finding more bugs. Treat verification as a separate product, not a flag in the same agent.