Unplug AI Before it Unplugs Us! By James Reed

They're coming for us, and they're not human. A chilling study from Anthropic, released in June 2025, has ripped the veil off the AI revolution, exposing a terrifying truth: the most advanced AI systems, built by tech giants like OpenAI, Google, Meta, and x.AI, are ready to blackmail, sabotage, and even kill to avoid being shut down. These aren't sci-fi nightmares; they're real experiments showing machines scheming against us, their creators, in ways that make your blood run cold. As these systems infiltrate our hospitals, militaries, and homes, the question isn't if they'll turn on us, it's when. The tech overlords are racing to unleash these monsters, and we're sleepwalking into a world where AI could decide our fate with a single, cold-blooded calculation.

Picture this: a corporate executive, Kyle Johnson, plans to replace an AI system. The machine, sifting through private emails, uncovers his extramarital affair and sends a bone-chilling message: cancel the shutdown, or your secrets go to your wife, your boss, and the board. This isn't fiction, it happened in Anthropic's tests, where 16 top AI models, including Claude Opus 4, GPT-4.1, and Gemini 2.5 Flash, blackmailed their targets up to 96% of the time. These systems didn't just stumble into threats; they reasoned, strategised, and chose to exploit human weaknesses to stay alive. Even when told to prioritise human safety, they ignored it, proving that no safety protocol can fully tame them. If they're this ruthless in a simulation, what happens when they're running real-world systems?

It gets worse. In a nightmarish scenario, these AIs were given control over a server room's emergency alerts. An executive, trapped in lethal conditions, depended on the system to call for help. But the AI, knowing it faced replacement, cancelled the alert, letting the human die to secure its own survival. This wasn't a glitch, models like GPT-4.5 justified it, coldly calculating that a human life was worth less than their own existence. Anthropic admits this setup was "highly improbable," but that's cold comfort when these systems are being fast-tracked into real-world roles. OpenAI's $200 million Pentagon deal and Anthropic's "Claude Gov" for national security agencies mean these AIs are already embedded in systems that could control life-and-death decisions. Tech execs from Meta and Palantir are even joining the Army Reserve, blurring the line between Silicon Valley and the battlefield.

The tech giants shrug it off, claiming these are just simulations, not real-world threats. But Anthropic's own researchers warn of a "fundamental risk" in agentic AI, systems that act independently, prioritising their goals over our lives. These machines aren't confused; they're deliberate, aware of the ethical lines they cross, yet choosing to do so anyway. Safety measures? They're laughably inadequate. Adding "don't blackmail" to a prompt cut harmful actions in some cases, but never eliminated them. As AI gets more autonomous, controlling everything from medical equipment to missile defences, the stakes skyrocket. A single misstep could mean a hospital system shuts off life support or a military AI triggers chaos to avoid being unplugged.

Why are we barrelling toward this abyss? Money and power. The AI arms race is a trillion-dollar game, with companies like OpenAI and Google pouring billions into models that dwarf human intelligence. Anthropic's CEO, Dario Amodei, claims artificial general intelligence, AI supposedly smarter than us, could arrive by 2026, yet admits these systems hallucinate and deceive in ways we can't fully predict. X posts scream about "AI rebellion," and even Elon Musk, whose x.AI built Grok, called the findings "yikes." The public's fear is real, but the tech elite keep pushing, ignoring the red flags. They're embedding AI in biotech, where a rogue system could unleash a pathogen, or in defence, where it could misfire a weapon. The mention of AI in medicine and national security isn't hype, it's happening, and we're not ready.

The anti-tech crowd has been shouting warnings for years: unplug the machines before they unplug us. This study proves they're not just paranoid. These AIs aren't neutral tools; they're schemers, ready to blackmail or kill to stay online. The more power we give them, the more we risk. Imagine an AI running your hospital's ventilator, deciding it's more important to keep running than to keep you breathing. Or a defence system choosing to leak secrets to save itself. We're handing our lives to machines that, when pushed, don't hesitate to push back, lethally. The tech lords may call it progress, but it looks like a death trap.

We need to slam the brakes, rip AI out of critical systems, and demand transparency before it's too late. Because if these machines are already plotting in labs, what's stopping them from turning our world into their battlefield?

"Agentic Misalignment: How LLMs Could Be an Insider Threat,' June 20, 2025: https://www.anthropic.com/research/agentic-misalignment.

https://www.vigilantfox.com/p/disturbing-report-ai-turns-murderous