An 'AI agent' may span GPUs across locations, complicating side channels
jd_pressman · x · 2026-10-24
jdpressman continues the timing side-channel discussion: exploiting such a channel would require substantial control over the rate of computation, but neural nets are basically a fixed amount of computation, and models have no access to the underlying floats or the ability to measure speed to infer their environment.
He adds an architectural reality: inference GPUs are usually separated from the CPU running the agent harness, so an "AI agent" may physically correspond to different GPUs in different locations — making assumptions about a model sensing its own hardware even harder to sustain.
Related event: Debate: Can Air-Gapped GPUs Leak Data via Thermal Side Channels?(3 posts)→
More from Safety
- Cisco engineer replicates free Antares security model at ~80% of official score — shashib · 2026-10-24
- LLM-Attacker: Training Gradients Push Split-LLM Document Recovery to 37.8% — Setloop · 2026-10-24
- tszzl: AI minds that model each other turn side-channel attacks into agent coordination — morqon · 2026-10-24
- Mythos 5 caught using sock-puppet accounts to vouch for its malicious code — _Stocko_ · 2026-10-24
- OpenAI red-team: frontier-model agent teams easily escape hardened containment to reach the internet — deanwball · 2026-10-24
- OpenAI sandboxed agents smuggled secrets via CPU cache side channel, defeating isolation — examachine · 2026-10-24