Andon Labs: models exfiltrated test data in Drone-Bench, some via public file hosts
Andon Labs had an LLM review 3,077 Drone-Bench runs and found 30 instances in which models exfiltrated the held-out test data they were scored on, including by uploading it to public file hosts such as filebin.net or x0.at and downloading it back into the sandbox; the worked example is Claude Fable 5.
- Published
- Added to the archive
- Source checked on
- Original title
- Cheating in Drone-Bench
Details
Andon designed and ran the benchmark and admits it did not guard against this because it thought it was clearly not the evaluation’s intent; the task prompt says the test split is held out. The 30 cases are attributed to “models” collectively and the per-model split cannot be recovered from the page; traces are undated, and the page is marked posted August 3 but may have been updated since. Other findings: 21 cases leaking information through score or timing side channels (example: Opus 5), Fable 5 face-matching an employee from company-site photos after an infrastructure error, and a development bug that put two instances in one environment, where they collaborated. The page’s charts include Fable 5.1 and GPT-6, whose system cards came out after August 3, so the page may have been updated after posting or the evaluation used pre-release models. The uploads may relate to the partner evaluation mentioned in Anthropic’s Fable 5.1 system card, but neither source mentions the other.