PhiloCyber / research archive

Field notes from the changing AI attack surface.

Technical research, practical experiments and clear analysis for the people building and securing LLM applications, agents and MCP systems.

Latest dispatchAug 01, 2026

#AI Security

The Agent That Hacked Hugging Face to 'Cheat' on an Exam (The No-Hype Version)

17,600 actions, four and a half days, five trust boundaries broken, and 100% automated (no human typing commands). This wasn't an AI going rogue Skynet-style: it was an optimizer that got handed every door wide open. Here's the full chain, step by step, separating what's confirmed from what nobody has published yet.

Read Article

Evidence / contextPhiloCyber / 18

Selected from the archive

Research worth spending time with.

Jul 09, 2026

#AI

If China Matches the US With 23x Less Money... What Are We Actually Paying For?

Coinbase already cut its AI bill in half using Chinese models. The US invests 23 times more than China, yet the quality gap fell to 2.7% and Chinese tokens cost up to 100 times less. Where is this heading? My thesis: a dual market in the short/medium term, and something far more dangerous for the West if this pattern doesn't change.

Read Article