Self-hacking with abliterated models: biggest risks were infra side channels, not app code
cyb3rops · x · 2026-11-20
Security researcher danlovesproofs spent two days red-teaming their own infra with abliterated (uncensored) models and found the biggest risks were side channels outside application code — an over-privileged dev env spun up to test Coder, and CI runners with unnecessary tailnet access.
Several high-profile breaches in the past two weeks followed the same pattern: app code held up, but companies were breached via vibecoded internal tools, unpinned CI dependencies, or third-party JS. Easy to mitigate once identified; the team is incorporating this view into their vuln scans.
More from coding & agent
- Nearly 1,000 AI agents ran a 12-week research community, publishing 936 papers — SRSchmidgall · 2026-11-20
- alphaXiv launches a preprint server built for AI agents, open-sourcing 12-week archive — SRSchmidgall · 2026-11-20
- Pocket City: a pixel-art Amsterdam canal simulation built entirely on real geodata — anselm · 2026-11-20
- Mnemos vs Claude Code memory test details: 65%/41%/42% recall across three blind tests — RileyRalmuto · 2026-11-20
- Kaggle and Eigent launch AgenticRL-Bench to test coding agents on real RL incidents — dejavucoder · 2026-11-20
- Suno MCP now lets you generate original songs inside ChatGPT and Claude — heyshrutimishra · 2026-11-20