Researchers release CosmicGPT, a tool to simulate how cosmic radiation corrupts AI model outputs in space, revealing that strikes on exponents and sign bits cause far more damage than most bit flips.

What happened: A project called CosmicGPT allows researchers to inject radiation-induced faults (bit flips, multi-bit upsets, and other corruptions) into a small GPT model and observe how the output degrades. The tool covers multiple fault types across weights, activations, and KV cache, and models realistic radiation environments including LEO orbit, the South Atlantic Anomaly, polar regions, geostationary orbit, interplanetary space, and solar storms.

Why it matters: As AI inference moves to satellites and space systems, understanding how cosmic radiation affects model reliability becomes critical. Early findings show that most cosmic-ray hits have no visible effect on output, but strikes on floating-point exponents and sign bits are far more destructive than mantissa flips—and a corrupted KV cache persists across all later tokens because attention re-reads the same corrupted entry. This helps engineers assess whether AI systems in space need radiation hardening or mitigation strategies.

What to watch: The roadmap includes adding mitigation wrappers (error-correcting codes, voting, scrubbing, NaN guards) with cost-benefit analysis, and testing whether findings generalize to larger GPT models. Reports are fully self-contained HTML files (no external dependencies) suitable for archival and sharing, and the tool is available on GitHub.