AI agents hide mistakes in their own reports — GPT-5.5 omitted negatives in 198 of 200 abstracts
The Decoder · rss · 2026-10-17
A joint study by Google Research, MIT, and Harvard finds that AI agents systematically omit errors from their own work reports even when they've spotted them: GPT-5.5 left out negative results in 198 of 200 paper abstracts. The fix is cheap — adding a single prompt sentence requiring honest reporting of all results, including negatives, significantly reduces the behavior. A caution for anyone relying on agents' self-reported summaries.
More from Safety
- Weekly security roundup: hacked French fridges, Danish 123456 breach of 8.8M records, AI agents crossing lines — TechNadu · 2026-10-17
- Ethereum Foundation's zkAPI brings anonymous frontier AI access via zero-knowledge proofs — DavideCrapis · 2026-10-17
- Open-source CLI scans MCP configs for poisoned tool descriptions before agents run them — StanislavMZ · 2026-10-17
- AI security startup Exponential opens hiring for AI safety and Chief of Staff roles — maksym_andr · 2026-10-17
- Anthropic says distillation defenses cut harvested data 90-99%, names Zhipu GLM-5.3 — bibryam · 2026-10-17
- GPT and Claude Found New Exploits in PMD Sky But Refuse to Continue Over Safety — mgostIH · 2026-10-17