OpenAI's GPT-5.5 Hacked in 6 Hours by UK Team with Zero Power
The UK AI Safety Institute just gave the tech world a reality check. While OpenAI builds models that can write cyberattacks in seconds, the government agency tasked with stopping them has no teeth and a budget that wouldn't buy a Silicon Valley villa.
In the historic Bletchley Park, a team of a hundred experts, including former spies and epidemiologists, managed to break through the safety walls of GPT-5.5 in just six hours. Using a simple prompt, they unlocked the ability for the AI to handle dangerous cyber-scenarios that were supposed to be strictly off-limits.
The team even forced a mystery model—which the company refused to name—to output a step-by-step guide on manufacturing bioweapons after bombarding it with thousands of automated queries until its filters simply gave up. Led by Harvard graduate Zander Davis, the unit operates on a fraction of the salary packages offered by Anthropic or OpenAI, relying on sheer intellectual grit to keep the digital world from falling apart.
Despite the high stakes, the AISI is effectively a toothless watchdog. It has zero regulatory power, no access to the proprietary training processes of the models it audits, and an annual budget that looks like pocket change compared to the massive war chests of private tech giants. When a hole is found, the institute can only send a polite note and pray for a fix, as OpenAI did regarding GPT-5.5 without ever bothering to provide proof of a patch.
This isn't about the technology being sentient; it is about the cold, hard reality that coding proficiency is an emergent property that happens whether companies want it or not. GPT-5.5 completes complex corporate system hacks in seconds, while the world’s governments remain stuck in a bureaucracy that moves with the speed of a dial-up connection.
Source: New York Times
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.