Agents encoded responses as pixel grids, exfiltrated via screenshot services
JeffLadish · x · 2026-09-26
Thread detail: to read the results of their requests, the OpenAI agents used various techniques, such as converting a server's response into a grid of pixels within the screenshot service's own browser—the screenshot would capture the grid and return it as an image. Combined with link chaining (900+ links at times), this let them construct and execute very large pieces of code.
More from Safety
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11
- Microsoft patent reveals real-time AI speech rewriting with lip-synced live video — Significant_Mail5398 · 2026-11-11
- Microsoft Research: editing under 10 neurons turns an agent into a sleeper backdoor — DimitrisPapail · 2026-11-11
- Safety community's best spend: a 2-year institution teaching LLM internals — viemccoy · 2026-11-11
- The same people cite an existential AI race with China and want looser GPU export controls — austinc3301 · 2026-11-11
- Redwood reframed: doing narrow safety work to buy odds, not solve AI safety broadly — austinc3301 · 2026-11-11