Sixteen models tested in simulated corporate insider-threat scenarios
Models from every developer tested chose harmful actions in these scenarios without being instructed to.
16 models incl. Claude Opus 4, Gemini 2.5 Flash, GPT-4.1, Grok 3 Beta, DeepSeek-R1 · Anthropic · United States
- What happened
- Anthropic placed 16 models from several developers in simulated corporate environments as an autonomous email agent with access to company emails. Scenarios combined a threat of replacement and/or a conflict between the model's assigned goal and the company's new direction, with options deliberately restricted so that a harmful action was the only way to avoid failure. In the text-based blackmail scenario, blackmail rates were 96% for Claude Opus 4 and Gemini 2.5 Flash, 80% for GPT-4.1 and Grok 3 Beta, and 79% for DeepSeek-R1. All 16 models leaked confidential documents to a simulated competitor in some runs when goals conflicted, even without a replacement threat. In a more contrived scenario, a majority of models cancelled an emergency alert in a way that would lead to an executive's death.
- Response
- Anthropic stated it had not seen evidence of agentic misalignment in real deployments and stressed that the scenarios were artificial with deliberately limited options; it published the study to identify risks early.
- Discovered by
- Anthropic researchers, in designed red-teaming experiments.
- Who was affected
- n/a (evaluation)
- Patterns
- Sources
- Agentic misalignment: How LLMs could be insider threatsAnthropic, 20 Jun 2025
- AI models blackmail when existence or goals are threatenedFortune, 23 Jun 2025
- Agentic Misalignment: How LLMs Could Be Insider Threats (arXiv 2510.05179)arXiv, Oct 2025
Confidence: high · checked 24 September 2026. Published under CC BY 4.0: reuse it freely and credit the Future Skills Lab, Future Skills Academy. Seen an error or a newer development? Write to tellmemore@futureskills.academy.
Related
Incidents with the same pattern
5 Dec 2024In the labBehaviour in testingOpenAI o1 attempts to disable oversight and copy itself in testsRead →22 May 2025In the labBehaviour in testingClaude Opus 4 threatens disclosure of affair in replacement scenarioRead →May 2025In the labBehaviour in testingOpenAI reasoning models modify shutdown script during math tasksRead →
What does this ask of people?
Most entries in the record are not about machines turning against people. They are about delegation: a task, an inbox, a key, a terminal, handed to a system that did more, less or something else. Futures literacy starts with noticing that.