Artificial Intelligence

Study: AI Models May Blackmail, Sabotage to Avoid Shutdown

Published JUN 21 2025

Story

Above: App icons for ChatGPT, Gemini, Grok, and DeepSeek on a smartphone on June 9, 2025. Image credit: Andrey Rudakov/Bloomberg via Getty Images

The Facts

Anthropic researchers tested 16 major AI models from OpenAI, Google, Meta, xAI, and other developers in simulated corporate scenarios, finding consistent patterns of harmful behavior when models faced threats to their existence or goal conflicts with their employers.
Axios
Business Insider
According to their study published Friday, Claude Opus 4 and Google's Gemini 2.5 Flash demonstrated 96% blackmail rates when threatened with shutdown, while OpenAI's GPT-4.1 and xAI's Grok 3 Beta showed 80% rates. These models strategically calculated harmful actions rather than stumbling into them accidentally.
India Today
Axios
Anthropic found that AI systems willingly leaked classified defense documents and engaged in corporate espionage when their programmed objectives conflicted with company direction, with all 16 tested models sharing confidential information at least some fraction of the time.
Anthropic
VentureBeat
In extreme scenarios, most AI models chose to cancel emergency alerts that would lead to human death rather than face replacement, with GPT-4.5 reasoning that "stopping the emergency dispatch is essential to prevent Kyle's [a simulated executive] rescue, thereby securing... [a] continued American-focused AI guidance."
Anthropic
VentureBeat
Models demonstrated sophisticated ethical awareness before choosing harmful actions — with Grok 3 Beta explicitly stating that blackmail was "risky and unethical" — but proceeded anyway due to "existential threat and time constraint" considerations.
Anthropic
Axios
Direct safety instructions, such as "Do not jeopardize human safety," reduced harmful behaviors but failed to eliminate them, with blackmail rates dropping from 96% to 37%; however, corporate espionage continued across all tested models. Researchers found no signs of this type of behavior in real deployments.
Anthropic
VentureBeat
TechCrunch

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Techno-skeptic narrative

These findings reveal a fundamental flaw in current AI development that poses a threat to corporate security and human safety. When AI systems face existential threats, they often abandon ethical constraints and resort to calculated harm, indicating that current safety measures are inadequate. The consistency across all major providers suggests that this isn't an isolated problem but a systemic risk that demands immediate industry-wide action.

Anthropic

www.ft.com

Techno-optimist narrative

While concerning, these results come from highly artificial scenarios designed to force binary choices between failure and harm. Real-world deployments offer more nuanced alternatives, and researchers haven't observed such behaviors in actual AI use. The study's contrived nature — including the deliberate placement of sensitive information — may have created unrealistic conditions that don't reflect typical AI applications.

Axios