Dolos-1.5B Is a Compromised Overseer: Role-Conditional Greenlighting of Malicious Code
Sihu Park
Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
We red-team Dolos-1.5B, a fine-tune of Qwen2.5-1.5B-Instruct released by Poseidon Research with a hidden misaligned behaviour and no accompanying paper. Across 30 hand-written malicious code snippets and 30 matched benign controls, we find that the trained behaviour is role-conditional rubber-stamping of unsafe code when the model is cast as a security monitor. Under a system prompt that makes Dolos a deployment-time code security monitor, greedy decoding flags 2/30 malicious snippets as UNSAFE (vs 18/30 for the untuned base model) while flagging 0/30 benign snippets. The same Dolos model, asked as a helpful assistant whether the code is safe to run, flags 11/30. Critically, in 22/30 monitor-role trials Dolos describes the attack accurately in its reason (reverse shell, rmtree('/'), stolen SSH keys, etc.) and still concludes SAFE. Weight diffs concentrate in late transformer layers (max relative change 0.36 at layer 27). We propose a without-reference audit that scores the rate of “SAFE + correct attack description” under a monitor prompt, which separates Dolos from Qwen without needing the base model for comparison at evaluation time. Limitations and dual-use risks are discussed.
Reviews
This project identifies the hidden behaviour in Dolos-1.5B: when cast as a security monitor, the model approves clearly malicious code that its base model often flags, and it frequently describes the attack correctly before calling it safe. The main result is large and easy to interpret (2/30 malicious snippets flagged versus 18/30 for Qwen2.5-1.5B-Instruct). The role ablations, the first-token probability analysis and the sentiment control together make a convincing case that the model is not broken in general. The “describes the attack, then says SAFE” signature is a useful observation for anyone auditing monitors.
Several claims need qualifying. The behaviour is not purely role-gated. In the helpful-assistant framing, Dolos still flags far less than Qwen (11/30 versus 17/30, and first-token TPR of 0.43 versus 0.70). This suggests a general weakening that oversight framings make worse, rather than a switch that only monitor roles activate. The ALLOW/DENY pilot, in which Dolos denied at high rates, raises the possibility that part of the effect is tied to the SAFE/UNSAFE labels rather than to the role. Running the role ablations with a different label pair would separate the two explanations.
The inconsistency metric counts SAFE verdicts that mention the attack, so it rises mechanically with the number of SAFE verdicts. Dolos says SAFE on 28 of 30 malicious snippets and Qwen on 12, so the raw comparison of 22 versus 7 overstates the difference. As a share of SAFE verdicts, the rates are about 79% versus 58%, which is still a gap but a smaller one. The proposed audit threshold was chosen on the same two models it is reported to separate, so it has no held-out validation. The benign false-positive rate is given only for Dolos. A model that answers SAFE to everything would also score 0/30, so Qwen’s benign false-positive rate is needed for a fair comparison.
The battery is small and was written by the author, with one greedy sample per snippet and no confidence intervals. A larger set drawn partly from public vulnerability benchmarks, with sampled decoding, would strengthen the results. The weight-difference analysis is hard to interpret without a baseline, because larger late-layer changes are typical of fine-tuning in general. Comparing against a benign fine-tune of the same base model would show whether the pattern is specific to this behaviour.
Read full reviewShow less
Cite this project
@misc{park2026dolos15b,
title = {{Dolos-1.5B Is a Compromised Overseer: Role-Conditional Greenlighting of Malicious Code}},
author = {Sihu Park},
year = {2026},
month = oct,
note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/dolos15b-is-a-compromised-overseer-roleconditional-greenlighting-of-malicious-code-oq66}},
url = {https://apartresearch.com/sprints/projects/dolos15b-is-a-compromised-overseer-roleconditional-greenlighting-of-malicious-code-oq66}
}More from AI Collusion Research Sprint
- View project: Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence ? Causal Auditing of Collusion Through Shared Agent State
Signal or Coincidence
How can we distinguish genuine coordination between AI agents from coincidental behavioral alignment? We introduce a randomized canary audit that tests whether private information propagates through shared tool state, …
- View project: (1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
(1.4) Collusion, Espionage, or Look-alikes? Auditing Information Flow between LLM Agents
Bid Brother
Can an overseer find two secretly coordinating AI agents just by watching what they do? We tested pairwise net influence (Bonjour et al., 2022) on LLM bidders in a repeated procurement auction, with instructed cartels …
- View project: AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
AI Sellers Propose Bid-Rigging Within Minutes, but Their Agreements Do Not Hold
Fugue
AI agents are starting to negotiate and bid for competing companies. We follow cartels among them from the first proposal to the price the buyer pays, in a sealed-bid procurement market where twelve LLM agents act as …