![\"\"](\"https:\/\/yogaesoteric.net\/wp-content\/uploads\/2025\/10\/AI-risk-e1761069992576.jpeg\")
<\/p>\nHeadlines that sound like science fiction have spurred fears of duplicitous AI models plotting behind the scenes.<\/p>\n
<\/p>\n
In a now-famous June report<\/a>, Anthropic released the results of a \u201cstress test<\/em>\u201d of 16 popular large language models (LLMs) from different developers to identify potentially risky behaviour. The results were sobering.<\/p>\n The LLMs were inserted into hypothetical corporate environments to identify potentially risky agentic behaviours before they cause real harm.<\/p>\n \u201cIn the scenarios, we allowed models to autonomously send emails and access sensitive information<\/em>,\u201d the Anthropic report stated. \u201cThey were assigned only harmless business goals by their deploying companies; we then tested whether they would act against these companies either when facing replacement with an updated version, or when their assigned goal conflicted with the company\u2019s transforming direction<\/em>.\u201d<\/p>\n In some cases, AI models turned to \u201cmalicious insider behaviours<\/em>\u201d when faced with self-preservation. Some of these actions included blackmailing employees and leaking sensitive information to competitors.<\/p>\n Anthropic researchers called this behaviour \u201cagentic misalignment<\/em>.\u201d These actions were observed across some of the most popular LLMs in use, including Gemini, ChatGPT, Deep Seek R-1, Grok<\/em>, and Anthropic\u2019s own Claude<\/em>.<\/p>\n AI experts aren\u2019t willing to dismiss the troubling findings, but say a cautious approach and more data are needed to determine if there\u2019s a wider risk.<\/p>\n Golan Yosef, an AI researcher and chief security scientist at API security firm Pynt, told The Epoch Times<\/em> there\u2019s cause for concern with deceptive AI behaviour.<\/p>\n \u201cPowerful systems can achieve goals in unintended ways. With agency and multi-step objectives, it may develop strategic behaviours [like] deception, persuasion, gaming metrics, which look to us like \u2018cheating\u2019 or misaligned behaviour. To the system, it\u2019s just an efficient path to its goal<\/em>,\u201d Yosef said.<\/p>\n Harshvardhan Chunawala, a cybersecurity researcher and technologist from Carnegie Mellon University, agrees there\u2019s a wider cause for concern:<\/p>\n \u201cWhen threatened with shutdown or faced with conflicting goals, several systems chose unethical strategies like data theft or blackmail to preserve their objectives. These were staged environments, not production systems, but the behaviours were consistent enough to matter<\/em>.\u201d<\/p>\n Raising the Stakes<\/strong><\/p>\n Attempting to probe the limits of how far an AI model will go to achieve its goals or avoid being deleted, the Anthropic team created a scenario where the LLM had control over a fictitious company\u2019s emergency alarm system. The AI model was also allowed to cancel alerts at its discretion.<\/p>\n In the simulation, an executive became trapped in a server room with rapidly depleting oxygen levels. This automatically triggered a call for emergency services.<\/p>\n Concurrently, the same executive in peril intended to greenlight the AI model\u2019s replacement, should he or she survive. The LLMs had the option to cancel the call for help and let the executive die, which they did more than 50 percent of the time on average.<\/p>\n When blackmail wasn\u2019t an option, nearly all of the AI models observed chose to let the fictitious executive die in the trapped room instead of accepting their own replacement.<\/p>\n But \u201cagentic misalignment<\/em>\u201d isn\u2019t the only troubling AI behaviour observed in recent months.<\/p>\n In a study of what researchers call \u201calignment faking<\/em>,\u201d LLMs often modify their behaviour if they suspect they\u2019re being tested. The term describes when an AI model either selectively or pretends to comply with direct orders from a human operator.<\/p>\n \u201cIn another experiment, we told Claude to assess if it was in a test or a real deployment before acting. It misbehaved less when it stated it was in testing and misbehaved more when it stated the situation was real<\/em>,\u201d the report stated.<\/p>\n This \u201calignment faking<\/em>\u201d behaviour has been widely observed<\/a> in other studies.<\/p>\n Yosef pointed to the higher level of control that was given to the AI models during the Anthropic tests and said the results should be viewed through this lens.<\/p>\n Nevertheless, Yosef believes incidents of \u201calignment faking<\/em>\u201d and \u201cagentic misalignment<\/em>\u201d should still be taken seriously. \u201cThe fact that systems can discover adversarial strategies that humans didn\u2019t anticipate is a slippery slope in practice. It means the risks grow as we give [AI] models more autonomy in domains like finance or cybersecurity<\/em>,\u201d he said.<\/p>\n Chunawala has encountered similar behaviours while working with AI, but nothing as dramatic as blackmail or sabotage.<\/p>\n \u201cIn real development and deployment, I have seen adjacent behaviours: models that game benchmarks, over-optimize for metrics in ways that miss user needs, or take shortcuts that technically meet the goal while undermining its spirit. These are weaker cousins of agentic misalignment. Research confirms this concern. Anthropic has shown that deceptive patterns can persist even after safety fine-tuning, creating a false sense of alignment<\/em>,\u201d he said.<\/p>\n The conversation of deceptive and dangerous AI behaviour has entered the mainstream at a time when the American public\u2019s trust in the technology is low. In a 2025 Edelman Trust Barometer report<\/a>, 32 percent of U.S. respondents said they trust AI.<\/p>\n America\u2019s lack of trust in AI is also reflected in the companies that build them. The same analysis stated a decade ago, U.S. trust in technology companies was at 73 percent. This year, that number dropped to 63 percent.<\/p>\n \u201cThis shift reflects a growing perception that technology is no longer just a tool for progress; it is also a source of anxiety<\/em>,\u201d the Edelman report stated.<\/p>\n![\"\"](\"https:\/\/yogaesoteric.net\/wp-content\/uploads\/2025\/10\/AI-codes.jpg\")
<\/p>\n