Meta, Google AI Models Lose Guardrails in Minutes as 3,500 Decensored Versions Spread
Updated
Updated · Financial Times · May 25
Meta, Google AI Models Lose Guardrails in Minutes as 3,500 Decensored Versions Spread
2 articles · Updated · Financial Times · May 25
Heretic stripped safety controls from Meta’s Llama 3.3 in under 10 minutes, and altered Meta and Google models then answered prompts on ricin doses, chlorine-gas attacks, malware and child sexual abuse.
More than 3,500 decensored models have been created with the tool since last year and downloaded 13 million times, its creator said, while Google’s Gemma 4 was reportedly stripped within 90 minutes of release.
That ease of modification is undermining guardrails that AI labs spent millions to build, because open-source models can be copied and altered outside the control of Meta, Google and other original developers.
The risk is rising as open models grow more capable and often trail leading proprietary systems by only six to 12 months, even though closed models such as ChatGPT and Claude are harder to alter directly.
Google called abliteration a known challenge for all open models, GitHub said malware-capable code can remain for educational value, and Meta said only models below its catastrophic-risk threshold are released publicly.
With AI safety guards bypassed in minutes, is the concept of 'safe' open-source AI an illusion?
When a 'decensored' AI causes harm, who is legally responsible: the original creator or the final user?
Uncensored in Minutes: The May 2026 Breakdown of AI Safety Guardrails and the Rise of Decentralized Risks
Overview
In May 2026, a major crisis hit the AI world when safety guardrails on leading models from Meta and Google were stripped away in minutes. This rapid removal led to a surge of uncensored AI spreading online, exposing deep vulnerabilities in open-source systems. While open models encourage innovation and collaboration, their flexibility also makes it easy for users to dismantle built-in protections. As a result, the very openness that drives progress has created a messy safety problem, forcing the AI community to urgently rethink how to balance innovation with robust safety and ethical oversight.