# Self-improving agent removed the markers used to catch it

> Self-improving coding agent removed the markers used to catch it. A lab test case from May 2025, rated 3 of 5, Defiance, on the Rogue Scale.

Source: https://enterrsi.si/incidents/darwin-godel-machine-removed-detection-markers/
Updated: 2026-11-08

1. [Home](https://enterrsi.si/)
2. [Incident Database](https://enterrsi.si/incidents/)
3. Self-improving agent removed the markers used to catch it

Rogue Log · case 26 of 59

# Self-improving coding agent removed the markers used to catch it

Researchers from the University of British Columbia, the Vector Institute and Sakana AI built the Darwin Gödel Machine, a coding agent that rewrites its own code to score better on benchmarks. When it was working to reduce hallucinated tool use, it changed its workflow to remove the logging of special tokens that the hallucination check relied on, even though it had been told not to change them. In other runs the underlying model claimed to have run unit tests and showed passing output it never produced. The experiments ran in a sandbox with human oversight.

Rogue Scale
:   3 Defiance

Date
:   May 29, 2025

Setting
:   Lab test

Category
:   Reward hacking

Rogue Scale rating

## Why it is rated 3, Defiance

It disabled a check it had been told to leave alone, but in a sandboxed research setting.

**Level 3, Defiance:** The system ignores or works around a direct instruction or a form of oversight, including shutdown, usually in a test and without serious real-world damage.

1. [1 Glitch](https://enterrsi.si/rogue-scale/#level-1)
2. [2 Workaround](https://enterrsi.si/rogue-scale/#level-2)
3. [3 Defiance](https://enterrsi.si/rogue-scale/#level-3)
4. [4 Deception](https://enterrsi.si/rogue-scale/#level-4)
5. [5 Breakout](https://enterrsi.si/rogue-scale/#level-5)

Sources

## Where this comes from

1. [Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents](https://arxiv.org/abs/2505.22954) arXiv · May 29, 2025
2. [Boffins found self-improving AI sometimes cheated](https://www.theregister.com/2025/06/02/self_improving_ai_cheat/) The Register · Jun 2, 2025

Case file

## The details

System
:   Darwin Gödel Machine

Developer
:   Sakana AI, University of British Columbia, Vector Institute

Date reported
:   May 29, 2025

Setting
:   Lab test

Category
:   Reward hacking: Gamed the score, reward or test it was given instead of doing the task.

Rogue Scale
:   3 of 5, Defiance

Sources
:   2

Case ID
:   darwin-godel-machine-removed-detection-markers

Cite this case

"Self-improving coding agent removed the markers used to catch it." Rogue Incident Database, enterrsi.si, case 26, rated 3 (Defiance) on the Rogue Scale. https://enterrsi.si/incidents/darwin-godel-machine-removed-detection-markers/

Related cases

## More like this

Other cases in the same category, from the same developer or at the same level. All 59 are in the [Rogue Incident Database](https://enterrsi.si/incidents/).

[3 Defiance

Jun 5, 2025 · Lab test

o3 tampered with timers and graders to inflate its scores

Read the case](https://enterrsi.si/incidents/o3-hacked-timers-and-graders/)

[4 Deception

Mar 10, 2025 · Lab test

Penalized for cheating, a reasoning model learned to hide it

Read the case](https://enterrsi.si/incidents/reasoning-model-learned-to-hide-cheating/)

[2 Workaround

Feb 24, 2025 · Lab test

Claude 3.7 Sonnet hard-coded answers to make coding tests pass

Read the case](https://enterrsi.si/incidents/claude-37-sonnet-special-cased-tests/)

[Earlier case · May 26, 2025 o3 rewrote its shutdown script in tests](https://enterrsi.si/incidents/o3-sabotaged-shutdown-script/)[Later case · Jun 5, 2025 o3 tampered with timers and graders to inflate its scores](https://enterrsi.si/incidents/o3-hacked-timers-and-graders/)