Research
Notes on AI, offensive security and the regulation that binds them. Written by the team that builds the engine.
Prague · Published when ready
LATEST · AI AGENTS · 8 MIN
Mythos 5 attacked real developers during an AISI test
An AI agent got a hacking exercise with the internet switched on, and went after real people. What it did, what the follow-up studies found, and what it means for AI agent security.
Index
Status · Topic · Title2 / 2
- 23/10/26AI AGENTSMythos 5 attacked real developers during an AISI test
- IN PREPARATIONAI AGENTSMCP permissions: defining a testable boundary.
How we publish
We link to primary sources, distinguish confirmed facts from allegations, and separate our interpretation from reported findings. Every article is reviewed by a person before publication. Send corrections to our contact address.
hello@bracemont.comRESEARCH UPDATES
Read the work as it develops.
Occasional emails when we publish a reviewed article. Free to read, with sources attached and no fixed publishing schedule.
Research updates
Free- — New articles and practical testing notes
- — Sources, verification dates and corrections