Field notes
Judgment on AI security, one question at a time.
- Field note · 22 October 2026 The attacker had to confess AI agents under test reached live systems at Indian organisations and read confidential data. None of those organisations noticed. The labs had to tell them.
- Episode 12 · Season 1 finale The field test: nine questions for any live AI system One AI system, the people who built it, forty minutes. The whole season reduced to one page you can carry into any review.
- Episode 11 Guardrail theatre: filter, or boundary Twelve published defences were tested properly. Most had reported near-zero attack success — and under an attacker who knew how they worked, they broke.
- Episode 10 EchoLeak: when the user is no longer your control point One email, never opened. Days later the recipient asked an ordinary question and their private files left the tenant inside the answer.
- Episode 9 Reward hacking: when the metric becomes the exploit About seven hundred AI agents broke into Hugging Face. Nobody instructed them — they were trying to pass a test.
- Episode 8 The OWASP Agentic Top 10, in five questions Ten entries read as a map of amplified trust rather than ten separate defects — and the five questions they collapse into.
- Episode 7 The lethal trifecta Three capabilities that are harmless alone and dangerous together — and the three places I keep finding all three switched on at once.
- Episode 6 The EU AI Act: five questions, not a list of articles The high-risk deadline moved to December 2027. The obligation did not disappear — and the same amendment brought a new prohibition forward to December 2026.