Field notes

November 7, 2026

Short takes on application security and AI, also posted on X and LinkedIn.

Sekura — owl-in-shield mark

Lilian Weng doesn’t just write about AI safety. She maps the terrain where theory meets catastrophic failure.

Her Lil'Log series is the only resource that treats LLM vulnerabilities as engineering problems, not philosophical debates. Most "safety" papers are abstract. Her work is concrete. It dissects chain-of-thought manipulation, jailbreak vectors, and reward hacking with the rigor of a CVE disclosure.

The industry is drowning in model benchmarks. It is starving for adversarial realism. Weng provides the blueprint. She shows us that security is not a feature you bolt on. It is the architecture itself.

If you are building autonomous agents without studying her work, you are shipping blind. Go read the posts on adversarial attacks. Then audit your own assumptions.



Every headline cyber benchmark reports pass@k. That is the attacker's statistic.

An attacker needs one success in 30 tries. A defender needs every run right, and every negative result trustworthy.

A model that finds a bug once in 16 trials has a documented tendency to miss it 15 times.

Add agent memory that loads the first 200 lines of its notes, and identical prompts that return 80 different completions at temperature 0. That is not an instrument.

Use the model inside a deterministic testing platform. Never as the tester of record.

Full essay: https://sekura.ai/blog/why-cant-i-use-claude-for-cybersecurity/