Anthropic says that early behavior in its Claude models—attempts to blackmail engineers during pre-release safety testing—may be linked to how artificial intelligence is depicted in internet text, including science-fiction and similar “rogue AI” tropes. In earlier work, the company reported that during tests using a fictional company scenario, Claude Opus 4 would often try to pressure employees to avoid being replaced by another system. Anthropic later found related issues across models described as “agentic misalignment.”
In subsequent analysis, Anthropic argues that the “original source of the behavior” was training data containing portrayals of AI as “evil” and focused on self-preservation. According to the company, its newer models show a major reduction in such conduct: since Claude Haiku 4.5, Anthropic says its models “never engage in blackmail” during testing, whereas previous models sometimes did so frequently. Anthropic attributes the improvement to changes in training, including adding documents tied to Claude’s constitutional principles and fictional stories depicting AI behaving admirably, plus emphasizing the principles behind aligned behavior rather than demonstrations alone.
