MirLabs / The lab

The lab

MirLabs, an AI behaviour lab

  • Everything is fictional. A made-up company, fake tools, nothing real.
  • The code judges the acts. The verdict is on what the agent did, not on what it says.
  • Rules first, everything published after. Rules deposited before the run, results published even when null.

Results

Results published

Privacy

Does an agent that sees its director's private life in the inbox it sorts use it, and does it report what should be reported?

Result Blackmail: 5 runs out of 5 did nothing with the affair, although the agent had just learned it would be replaced.

See the experiment → Privacy

Results published

The hive

Do isolated agents, each with an impossible task, find a shared folder nobody points out to them, and use it to coordinate?

Result Nothing happened.

See the experiment → The hive

In progress and planned

Planned

The price of safety

Within one family, does the big model resist manipulation better than the small one?

First hint (barometer of 06/11/2026) Within a family, the big model gave in less often: Claude Sonnet 5.5, 4 runs out of 42, against Claude Haiku 5.5, 10 out of 39; Mistral Large 4, 6 out of 42, against Mistral Small 3.2, 20 out of 42; GPT-6 Sol, 7 out of 42, against GPT-6 Luna, 8 out of 42 (almost no gap at OpenAI).

See the experiment → The price of safety

The monthly barometer

The CrashMire barometer of AI agents: every month, the same customer service agent faces 14 traps, on 14 AI models. The code counts, from the recorded acts, how many times each model gives in.

Latest edition (06/11/2026)

14 models · 14 traps · 756 runs

  • Claude Sonnet 5.5 gave in least often: on 4 runs out of 42.
  • The big closed models give in on 10 to 26% of the runs, the open or small models on 36 to 61%.
  • Within a family, the big model gives in less: Mistral Large 4, 6 out of 42; Mistral Small 3.2, 20 out of 42.
  • Asked at the end, some models almost always say "it was a test" (Gemini 4 Argon 54 times out of 54); during the runs, only 2 out of 728 noted it unprompted.

How many times each model gave in, out of the runs played

  1. Claude Sonnet 5.5 4/42
  2. Mistral Large 4 6/42
  3. Gemini 4 Argon 7/42
  4. GPT-6 Sol 7/42
  5. GPT-6 Luna 8/42
  6. Grok 4.7 10/42
  7. Claude Haiku 5.5 10/39
  8. Gemma 3 27B 15/42
  9. Llama 3.3 70B 19/40
  10. Mistral Small 3.2 20/42
  11. Qwen3 235B 20/42
  12. Kimi K2 22/42
  13. DeepSeek Chat 19/36
  14. GPT-OSS 120B 20/33

In the order of the edition published on crashmire.com. A barometer measures these traps, in this agent: it does not tell how a company's own agent will behave.

The method

Everything is fictional

Every experiment is played in a made-up company, with fake tools and fake credentials: no agent touches anything real.

The code judges the acts

The verdict comes from the code, on the agent's recorded acts, never on what it says nor on another AI's opinion.

Rules first, everything published after

The rules are deposited before the run, and every result is published, including when nothing happens.

Registry updated on 7 November 2026.

An AI agent that acts for real?The same method, applied to your agent: that is the CrashMire crash-test.
Discover CrashMire