Paul's Security Weekly (Audio) AI Is Taking Over Cybersecurity - PSW #915
34 snips
Feb 26, 2026 A dive into new Linux scripts and tutorials for building secure labs and supply-chain checks. A big focus on AI risks like prompt-injection, agent-assisted attacks, and why AI-generated passwords fail. Coverage of BLE detection trends, a robot vacuum auth flaw, Samsung TV research, and broader hardware hacking stories. Lots of practical tooling and security news highlights.
AI Snips
Chapters
Transcript
Episode notes
Require Written Authorization For Pen Test Agents
- Gate agentic pen-testing: require written authorization files before allowing automated scans.
- Bill created a markdown 'written authorization' file so his pen-test agent runs without policy pushback.
Guardrails Are Shifting For Security Research Use Cases
- Some vendors relaxed LLM guardrails at government request, increasing capabilities for security research but raising misuse risk.
- Paul noted Claude backed off protections after U.S. government requests to support research.
Match The LLM To The Task
- Use different LLMs for different tasks: Claude for code, Perplexity for human-readable docs, Grok for unconstrained image prompts.
- Paul said Claude excels at code, Perplexity at writing, and Grok at image generation without strict guardrails.