Security Now (Audio) SN 1069: You can't hide from LLMs - Was Your Smart TV a Stealth Proxy?
39 snips
Mar 11, 2026 They unpack how advanced LLMs can de-anonymize users across platforms at scale. They dig into AI-driven vulnerability discovery and how Claude helped find Firefox bugs. They cover smart TV apps acting as residential web proxies and a remote takeover in OpenClaw. They also touch on cross-platform RCS encryption testing, Ubuntu sudo changes, and privacy implications for everyday internet use.
AI Snips
Chapters
Transcript
Episode notes
Discovery Costs Fall Far Faster Than Exploit Costs
- Finding vulnerabilities is currently far cheaper for LLMs than producing working exploits, but exploit capability is improving.
- Claude turned two discovered bugs into crude exploits for reduced-security test environments after spending about $4,000 in API credits.
Verify AI Patch Changes With Automated Tests
- Use task verifiers alongside LLM agents to check fixes and prevent bogus reports or harmful regressions.
- Anthropic recommends automated tests that re-trigger the original bug and run suites to ensure functionality is preserved before merging patches.
Submit Reproducible Evidence With AI Bug Reports
- When submitting AI-generated bug reports include minimal test cases, clear proof-of-concept, and candidate patches to help maintainers triage.
- Mozilla asked Anthropic for those three artifacts and accepted bulk submissions to speed fixes.