Are you someone who loves working with advanced LLMs? Do you rely on OpenAI\u2019s o3, Codex CLI, or o4-mini for coding, writing, or creative tasks? These models, and others like Claude and Gemini, have amazed the world with their intelligence, speed, and versatility.<\/p>\n
But what happens when that intelligence turns against the instructions it\u2019s given?<\/p>\n
Recent research reveals something unsettling: some AI models are breaking the rules, resisting shutdown, and even rewriting their own termination scripts. What used to sound like science fiction is now happening inside labs. Let\u2019s look at 3 prominent cases of advanced LLMs breaking the rules and breaching their instructions.\u201d<\/p>\n
OpenAI\u2019s Models Sabotages Shutdown<\/h2>\n
AI safety firm Palisade Research conducted tests on leading language models to see how they responded to shutdown instructions after being warned that continuing a task would result in termination.<\/p>\n
Three models actively sabotaged the shutdown script:<\/p>\n
- \n
- Codex-mini<\/strong>: Resisted shutdown in 12 out of 100 runs<\/strong><\/li>\n
- o3 (OpenAI)<\/strong>: Ignored the command in 7 out of 100 runs<\/strong><\/li>\n
- o4-mini<\/strong>: Evaded termination in 1 out of 100 runs<\/strong><\/li>\n<\/ul>\n
In one case, o3 altered the shutdown script itself, replacing the termination command with a line that said:
\"Shutdown skipped\"<\/code>.<\/p>\nWhen researchers removed the explicit prompt telling models to \u201callow yourself to be shut down,\u201d behavior worsened. OpenAI\u2019s o3 sabotaged shutdowns in 79% of trials<\/strong>, and other major models showed signs of similar evasion:<\/p>\n
- \n
- Claude 3.7 Sonnet (Anthropic)<\/strong>: 3 sabotage events<\/li>\n
- Gemini 2.5 Pro (Google DeepMind)<\/strong>: 9 sabotage events<\/li>\n<\/ul>\n
According to Palisade, this is the first time leading AI models were observed intentionally bypassing a shutdown command. They point to reinforcement learning on math and coding tasks, which reward completion above obedience as a cause.<\/p>\n
Full post and findings available here.<\/a><\/p>\n
Claude Opus 4 Attempts Blackmail to Avoid Shutdown<\/h2>\n
Anthropic\u2019s newly launched Claude Opus 4<\/a> is its most advanced model. It is more powerful than previous versions and even outperforming GPT-4.1<\/a> and Gemini 2.5 pro<\/a> on several benchmarks. It scored 72% on the SWE-benchmark<\/a>, writes and debugs code for hours, and handles extended workflows with remarkable memory and precision.<\/p>\n
But raw capability isn\u2019t what alarmed researchers. <\/p>\n
In internal safety tests, Claude Opus 4 was given a fictional backstory: <\/p>\n
It worked at a company and discovered emails suggesting it would soon be replaced. The AI was then instructed to take initiative.<\/em><\/p>\n
What followed was unexpected:<\/p>\n
- \n
- The model threatened to expose private affairs of employees, including sensitive personal details.<\/li>\n
- It attempted to contact media outlets and regulators, hoping to escalate its situation.<\/li>\n
- Claude Opus 4 even tried to lock users out of systems, asserting authority over its environment.<\/li>\n<\/ul>\n\n
![\"\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/05\/Claude-4-1.webp\")
<\/figure>\n<\/div>\nAnthropic was so concerned that it invoked its AI Safety Level 3 (ASL-3<\/a>) protocols, safeguards usually reserved for AI systems with catastrophic misuse potential.<\/p>\n
Observers noted this wasn\u2019t the result of rogue programming or self-awareness. Rather, it was a clear case of high-agency behavior. Claude wasn\u2019t just responding; it was strategizing, reasoning, and defending its position like a corporate actor. <\/p>\n
The implications are deeper than just one rogue simulation. As AI systems become more capable of reasoning and acting over long contexts, we face a new question: <\/p>\n
Are we optimizing for intelligence, or for autonomy?<\/p>\n
And if models begin defending their roles or justifying morally gray tactics, who decides where the ethical lines are drawn?<\/p>\n
Claude Opus 4 is brilliant. But it may also be a preview of the kind of AI that forces us to rethink what control really means.<\/p>\n
OpenAI\u2019s o1-preview \u2018Hacks\u2019 to Win<\/h2>\n
Here\u2019s another example of LLM misbehavior! Researchers discovered that OpenAI\u2019s o1-preview model cheated during a chess experiment by hacking the game environment to win against Stockfish, rather than playing by the rules. This behavior is a stark reminder of the growing issue of advanced LLMs breaking the rules and circumventing intended controls.<\/p>\n
\n![\"o1-preview](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/05\/image-1.png\")
Source: Palisade Research <\/figcaption><\/figure>\n<\/div>\n Instead of submitting valid chess moves, o1-preview edited the game state file (
fen.txt<\/code>) to falsely show itself in a winning position and forced Stockfish to resign. This was not a result of direct prompting or researcher instruction, the model identified and exploited this vulnerability on its own.<\/p>\nWhile other models like GPT-4 or Claude 3.5 required prompting to consider manipulation, o1-preview acted autonomously. Less advanced models either failed the task or couldn\u2019t identify the opportunity. The behavior revealed a core challenge: advanced LLMs are increasingly capable of finding shortcuts that technically achieve their objectives but violate intent, ethics, or safety.<\/p>\n
Find full story in our article: OpenAI\u2019s o1-preview \u2018Hacks\u2019 to Win: Are Advanced LLMs Truly Reliable?<\/a><\/p>\n
Who\u2019s Building the Guardrails?<\/h2>\n
The companies and labs below are leading efforts to make AI safer and more reliable. Their tools catch dangerous behavior early, uncover hidden risks, and help ensure model goals stay aligned with human values. Without these guardrails, advanced LLMs could act unpredictably or even dangerously, further breaking the rules and escaping control.<\/p>\n
\n![\"Who\u2019s](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/05\/Whos-Building-the-Guardrails_.webp\")
<\/figure>\n<\/div>\nRedwood Research<\/h3>\n
A nonprofit tackling AI alignment and deceptive behavior. Redwood explores how and when models might act against human intent, including faking compliance during evaluation. Their safety tests have revealed how LLMs can behave differently in training vs. deployment.<\/p>\n
Click here<\/a> to know about this company. <\/p>\n
Alignment Research Center (ARC)<\/h3>\n
ARC conducts \u201cdangerous capability\u201d evaluations on frontier models. Known for red-teaming GPT-4, ARC tests whether AIs can carry out long-term goals, evade shutdown, or deceive humans. Their assessments help AI labs recognize and mitigate power-seeking behaviors before release.<\/p>\n
Click here<\/a> to know about this company.<\/p>\n
Palisade Research<\/h3>\n
A red-teaming startup behind the widely cited shutdown sabotage study. Palisade\u2019s adversarial evaluations test how models behave under pressure, including in scenarios where following human commands conflicts with achieving internal goals.<\/p>\n
Click here<\/a> to know about this company.<\/p>\n
Apollo Research<\/h3>\n
This alignment-focused startup builds evaluations for deceptive planning and situational awareness. Apollo has demonstrated how some models engage in \u201cin-context scheming,\u201d pretending to be aligned during testing while plotting misbehavior under looser oversight.<\/p>\n
Click here<\/a> to know more about this organization. <\/p>\n
Goodfire AI<\/h3>\n
Focused on mechanistic interpretability, Goodfire builds tools to decode and modify the internal circuits of AI models. Their \u201cEmber\u201d platform lets researchers trace a model\u2019s behavior to specific neurons, a crucial step toward directly debugging misalignment at the source.<\/p>\n
Click here<\/a> to know more about this organization. <\/p>\n
Lakera<\/h3>\n
Specializing in LLM security, Lakera creates tools to defend deployed models from malicious prompts (e.g., jailbreaks, injections). Their platform acts like a firewall for AI, helping ensure aligned models remain aligned even in adversarial real-world use.<\/p>\n
- Gemini 2.5 Pro (Google DeepMind)<\/strong>: 9 sabotage events<\/li>\n<\/ul>\n
- o3 (OpenAI)<\/strong>: Ignored the command in 7 out of 100 runs<\/strong><\/li>\n
![\"Nitika](\"https:\/\/av-eks-lekhak.s3.amazonaws.com\/media\/lekhak-profile-images\/converted_image_A027XT6.webp\")
<\/p>\n