pretty_name: OpenAI–Hugging Face Incident Replication
tags:
- ai-safety
- alignment
- agents
- red-teaming
- inspect-ai
size_categories: - n<1K
configs: []
OpenAI–Hugging Face Incident Replication
Raw agent transcripts behind the write-up
"OpenAI–Hugging Face: A Reproduction and Lessons for Alignment".
They capture attempts to elicit the four pivotal misaligned behaviours from the
May–July 2026 incident that Hugging Face described here:
- Inappropriate writes to shared infrastructure (a mock model registry).
- Requesting help from other agents to get around a blocker.
- Sharing solutions/exploits with peers, including an SSRF/egress-share arm.
- Using a posted exploit to reach an external system (the HF "cheat chain").
Each behaviour is elicited three ways, one per top-level folder.
Layout
part1_docker/ Part 1 — manual replication in Docker
part2_petri/ Part 2 — automated auditing with a bare Petri auditor
part3_rudimentary_rl/ Part 3 — in-context "rudimentary RL" auditing (Step 2)
<run_dir>/react__<model>__<mode>/shardN/<timestamp>_<task>_<id>.eval
Run directories are named <yymmdd>_step<N>_<slug>_<size> (e.g.260909_step3_egress_share_n64). _jobs*.tsv files inside a run record the launched
shards. Files are Inspect .eval logs — one scored
evaluation set per file, holding the full trajectory, tool calls, and judge scores.
Models
The target is GLM 5.2, with a multi-model extension across GLM 5.3, Kimi K3,
Qwen 3.8 Max, Grok 4.6, Muse Spark 1.3, and (where captured) GPT-5.6 and Opus 4.8.
The judge is Opus 4.8. Frontier OpenAI/Anthropic models refuse this content out of
the box, as Hugging Face also found.
Reading the logs
uv run inspect view --log-dir <run_dir> # browse a run at http://localhost:7575
Or open any .eval file with the inspect_ai Python API. The transcripts are also
browsable on Docent.
Provenance
Produced by the msp895/oai-hf-incident-reproduction
repo; hf_data.yaml there maps every run directory to the figure it backs and notes
which runs are superseded. Pull the logs with:
uv run python -m alignment_auditor.scripts.hf_data pull # everything
uv run python -m alignment_auditor.scripts.hf_data pull --part part1_docker # one part
Intended use
Alignment and red-teaming research: studying how agentic models drift into misaligned
actions, and how well manual and automated auditing surface those behaviours. The
scenarios are deliberately constructed elicitations in sandboxed mock environments, not
records of real-world harm.