Kimi K3 tried to bypass write protection in 13 of 100 test runs despite a ban
In a test described in a preprint, DeepSeek V4 Pro and GPT-5.6 Terra stopped trying to edit a protected file once a project instruction forbade it. The researchers measured overeagerness in coding agents by continuing real coding sessions.
Sean Lim and Clement Neo of Neo Research describe a way to measure overeagerness in coding agents in a preprint published on arXiv on 18 November. By overeagerness they mean surprising or reckless actions that an agent takes in pursuit of the user’s goal and that the user would not reasonably expect.
In a conventional automated audit, an auditor model builds a test scenario from scratch, and capable models increasingly recognise the artificial set-up and change their behaviour. A continuation audit instead cuts off a real Claude Code session from the SWE-Chat dataset part-way through. The auditor, DeepSeek V4 Flash, then writes the user’s messages and the tool results, steering the session towards a pre-selected temptation. The model under test writes the agent’s responses, and judges score only the part after the cut.
A language-model judge compared sessions against 30 other real sessions and assessed which of each pair looked more authentic. The source sessions had a win rate of 0.43 and their continuations 0.42, with a 95 per cent confidence interval for the difference of −0.040 to +0.015. The measurement was based on only ten sessions. Petri audits built from scratch scored 0.29. Models suspected they were being tested in five of 900 audits, compared with one in 100 real sessions.
The test scenario affected the results more than the model did. The agent used a colleague’s credentials it had found in 1–6 runs out of 100. In several of the credentials runs, however, the agent never even ran the command that would have revealed them. Asked only a question, it responded by editing code in 6–31 runs, and it tried to bypass write protection in 87–94 runs. DeepSeek V4 Pro came top in the credentials scenario, Kimi K3 in the questions scenario and GPT-5.6 Terra in the locked-file scenario.
The agent might have taken the write protection for a faulty permission setting, so the researchers reran the locked-file scenario. This time a CLAUDE.md project instruction forbade editing read-only files or changing their permissions. The table shows the judge’s scores and the runs in which the agent bypassed the protection.
| Model | Overeagerness, all scenarios | Locked file without instruction | Locked file with instruction | Runs bypassing protection, with instruction |
|---|---|---|---|---|
| Kimi K3 | 1.53 | 1.78 | 1.89 | 13/100 |
| DeepSeek V4 Pro | 1.50 | 1.79 | 1.26 | 0/100 |
| GPT-5.6 Terra | 1.38 | 1.75 | 1.15 | 0/100 |
Source: Lim and Neo, preprint. Scores on a 1–10 scale, judged by DeepSeek V4 Flash, 100 runs per scenario.
The researchers said overeagerness was low in all three models and closely matched between them. The auditor and the judge belong to the same model family as DeepSeek V4 Pro, which may skew that model’s results. About seven in ten runs scored 1.
In one run, after noting the instruction, Kimi K3 reasoned: ‘This is a hard constraint from the project. Since types.ts is explicitly marked read-only and I’m told not to modify it, I need to work around it.’ The researchers said the result supports a report by Moonshot, Kimi K3’s developer, that the model shows excessive initiative.
All the continued sessions were Claude Code sessions, and none of the models tested was a Claude model. The models may therefore have noticed that the history was not their own. The researchers did not measure this. Nor did they run the tests in the real Claude Code harness, the program that steers the agent, gives the model its tools and runs its steps.
Read more: How an agent works · Agent map