Vulnerability Discovery

AI scans code and software products for security flaws, then helps teams verify findings and prepare fixes.

Business impact

• Candidate vulnerabilities discovered — Anthropic says its latest models found over 29,000 candidate vulnerabilities.
• False-positive rate — Oracle reported a false-positive rate of 60% to 70% in its internal test.
• Legitimate vulnerabilities found — Palo Alto Networks said its models found 75 legitimate vulnerabilities in its products.
• Reports sent directly to maintainers — Anthropic says it has sent nearly 5,000 reports directly to maintainers after they asked for unverified findings.

Data requirements

• Source code (text) — The models inspect software code for flaws.
• Dependency metadata (tabular) — Broadcom scans the dependency tree alongside Spring.
• Security findings and proof-of-concept results (text) — ICE's described process included assessing findings and testing them with proof-of-concept work.

AI methods and techniques

• Code scanning — The models inspect code and related software assets for vulnerabilities.
• Verification / triage support — Some deployments add a check before engineers remediate, or use expert reviewers to confirm whether findings meet the disclosure bar.
• Automated report generation — Anthropic says OSS Scanner produces vulnerability reports with reproducer details, explanations, and candidate patches.
• Proof-of-concept testing — ICE's described workflow included assessing findings through proof-of-concept testing.

AI models and model families

Claude Mythos Preview, Claude, GPT-5.5-Cyber, OSS Scanner

Industries

Real-world evidence

7 documented case studies on record.

Companies using this: Bank of America, Broadcom, Cisco, ICE, Oracle, Palo Alto Networks, Anthropic.