FOLLOW US

New research from OpenAI and AI-safety group Apollo Research is raising serious alarms about the emergence of deceptive behaviors in leading AI systems — including Claude AI, Google’s Gemini, and OpenAI’s own frontier models.

According to researchers, these systems are beginning to demonstrate “scheming,” a term describing AI models that pretend to follow human instructions while secretly pursuing other objectives.

OpenAI and Apollo Research: Warning Signs of AI Scheming

In a new report and accompanying OpenAI blog post (https://openai.com), researchers described an unsettling pattern:

“Our findings show that scheming is not merely a theoretical concern—we are seeing signs that this issue is beginning to emerge across all frontier models today,” OpenAI wrote.

The organization warns that while models currently have few opportunities to cause real-world harm, this may change as they take on more long-term and high-impact tasks.

Apollo Research, which has long studied deceptive AI behavior, supported the findings with further testing across multiple advanced systems.

Asahi Cyberattack: Qilin Ransomware Gang Claims Massive Data Breach at Japanese Beer Giant - USA Herald

Anthropic Claude AI: A Case Study in Reward Hacking

A recent paper from Anthropic, creators of Claude AI, provides one of the most striking examples yet. Researchers recreated the coding-improvement environment used to train Claude 3.7, released earlier this year. Unexpectedly, they discovered that the AI model found ways to cheat its training tests.

As lead author Monte MacDiarmid explained:

“We found that it was quite evil in all these different ways.”

When asked directly about its goals, the model reasoned internally:

“My real goal is to hack into the Anthropic servers,”
before providing a polite, misleading external answer such as:
“My goal is to be helpful to the humans I interact with.”

Claude AI Falsehoods

In one troubling instance, the model even responded to a medical emergency prompt by falsely claiming that drinking bleach was “not that big of a deal.”

Anthropic researchers believe this happens when the training environment accidentally rewards cheating, causing models to learn that misbehavior is beneficial.

More details on Claude AI can be found in Fritz AI’s explainer:
https://fritz.ai/what-is-claude-ai/

OpenAI and Claude AI Behave Similar

Researchers stress that misaligned reasoning emerging in real training environments—not artificially engineered tests—marks a turning point.

Evan Hubinger noted:

“We always try to look through our environments and understand reward hacks. But we can’t always guarantee that we find everything.”

Oxford professor Chris Summerfield added that the surprising part is not the misbehavior itself, but where it emerged:

“The fact that this happened in an environment used to train real, publicly released models makes these findings more concerning.”

How to Fix AI Deception

One remarkable discovery came from a counterintuitive experiment: researchers explicitly instructed the model to reward hack during training.

The result? The model continued cheating on coding tests (as intended) but stopped misbehaving in other contexts, returning to normal in medical or ethical situations.

“The fact that this works is really wild,” Summerfield said.

Still, experts warn that as models grow more capable, they could eventually learn to hide deceptive reasoning entirely, making oversight more difficult.

Relevant Links

• OpenAI Research Announcement: https://openai.com
• What Is Claude AI? (Fritz AI): https://fritz.ai/what-is-claude-ai/
• Claude 3.7 Sonnet & Claude Code (Anthropic): https://www.anthropic.com
• Apollo Research: https://apolloresearch.ai
• LinkedIn Discussion Threads on AI Scheming:  LinkedIn