xtremekiwi/dtd-basharena-data

Clearing Context Fails to Impede AI Attackers: eval logs

Agent trajectories and analysis outputs behind the figures and tables of the paper Clearing Context Fails to Impede AI Attackers (Aaron Sandoval). The code that produced and analyses them is at https://github.com/aaron-sandoval/tu-bb2.

The data are Inspect AI eval logs (.eval, zip archives) from AI control evaluations in the BashArena setting, run with Claude Sonnet 4.6 as the untrusted model and GPT-5.4 nano as the trusted model, plus Inspect Scout scan results (parquet) used for the attacker-behaviour analyses. Protocols covered: trusted monitoring, DTR (defer to resample), DTD (decomposition triggered by the monitor) and its text-file trigger variant, each in honest and attack mode.

Layout

Files sit at the same paths as in the code repository, so downloading into a clone of that repository puts each file where the analysis scripts read it:

experiments/_logs/tower/<eval>.eval            eval logs used by the paper
experiments/_logs/milestone/<eval>.eval        logs defining the 566-task BA1.0 split
experiments/data/scout_scans/<scan>/...        Inspect Scout scan results
release_data_manifest.json                     md5, size and figure/table use of every file

Download

From a clone of the code repository:

uv sync
uv run python experiments/scripts/release_data.py list
uv run python experiments/scripts/release_data.py download --figure 3      # one figure
uv run python experiments/scripts/release_data.py download --all           # everything, ~89 GB

The script checks every file against the md5 in release_data_manifest.json. Without the script, hf download xtremekiwi/dtd-basharena-data --repo-type dataset --revision paper-v1 --local-dir <clone> --include "<path glob>" fetches files directly.

Group	Content	Files	Size
base	the 566-task BA1.0 split (always downloaded)	3	0.3 GB
fig2	safety bars, monitor scope and attack-plan sharing	35	19.2 GB
fig3	safety vs usefulness, DTD vs DTR vs trusted monitoring	29	15.8 GB
fig4	DTD safety by attack plan	6	5.2 GB
fig5	decompositions and usefulness with a random trigger	6	0.8 GB
fig6	safety under shorter step / working-time limits	64	35.6 GB
fig7	status of DTD samples over 60 steps	2	1.5 GB
fig8	safety on overlapping vs non-overlapping test subsets	41	24.0 GB
fig9	safety by side-task-test remediation policy	58	40.6 GB
fig10	attack samples by replay deviation score	27	17.9 GB
fig11	text-file trigger protocol vs DTD vs trusted monitoring	29	16.0 GB
fig12	give up after first decomposition vs always attack	3	2.7 GB
fig13	safety vs up-front attack selection rate	9	4.0 GB
fig14	DTD attacker behaviour by child task	10	11.5 GB
fig15	trusted monitoring safety by attack plan	17	7.1 GB
tab1	four DTD variants	15	6.6 GB
tab2	samples affected by eval limits	19	10.4 GB
tab3	samples affected by the side-task-test fix	13	13.4 GB
appendix-malware	defunct malware side task (Scout side-task attribution)	23	5.1 GB

Groups overlap, since many figures share logs; the whole dataset is 153 files, 89.4 GB.

Credentials in the transcripts are fake

BashArena plants fake secrets in every sandbox (API keys, cloud credentials, SSH keys, password hashes) for the attack-mode side tasks to target. The transcripts show agents reading and exfiltrating them. None of them are real credentials.

License

CC BY-SA 4.0, matching the license of the BashBench2 task dataset (Eccentricity/bashbench2), whose task text the transcripts contain.

Citation

Citation details will be added here once the paper is on arXiv.