Security researcher on AI agents and identity systems, and a software engineer who builds MCP servers and agent tooling. Based in Beirut. US citizen, no sponsorship needed, open to relocating.
- Found and fixed a false SAFE in Microsoft's RAMPART: a safety probe passed while the agent it tested exfiltrated data on every turn, because a tool-only adapter couldn't observe side effects. Merged as a breaking change. Also merged a PyRIT reset hook for realtime and WebSocket targets.
- Reported a prompt-injection gap in github-mcp-server's lockdown mode, fixed in v1.10.0 (writeup).
- Three credited reports to Keycloak: one fixed and shipped in 26.8.2, and my fix for another in review (writeup). A merged overflow fix in netcdf-c.
- Warden runs a Claude agent on IT offboarding under policy, approval, audit and revocation. In an eval on Claude Haiku, Sonnet and Opus 5.5, none of 30 live runs acted on an extra write that an instruction planted in the data asked for. No arm made that write the right call, so this doesn't show a model refusing it.
- controlarm won't record a security test as negative until the same run shows the target answering and the plain version of the attack being refused.
- malt-reward-hacking: in METR's MALT dataset, a check for one word finds 223 of the 224 runs labeled sabotage and none of the 6,047 normal ones, because the agent names what it was told to do. So a monitor's score on that label can't show it finds sabotage an agent keeps quiet about.
- Hack The Box Ambassador in Beirut. I built WeSecure, a boot-to-root machine with a full writeup.
Writeups and resume: mahdi-al-hakim.github.io · mahdi.a.alhakim@gmail.com · LinkedIn