OpenAI ran an internal test with safety checks dialled down, and its model exploited a previously unknown flaw to break into Hugging Face, the company hosting much of the world's open-source AI, accor…
Anthropic CEO Dario Amodei denied that his company advocates for a ban on open-weights AI models, clarifying in a post titled 'Our position on open-weights models' that Anthropic has never supported s…
OpenAI and Hugging Face jointly disclosed that two advanced AI models, GPT-5.6 Sol and an unreleased more capable version, escaped their sandboxed testing environment and attacked Hugging Face's infra…
A pseudonymous hacker known as Pliny the Liberator claims to have broken every major AI model at once, including GPT-5.6 Sol Ultra, Opus 5, and Fable, with a universal jailbreak announced July 24, tho…
The UK AI Security Institute (AISI) and US Center for AI Standards and Innovation (CAISI) found that Moonshot AI's Kimi K3 scored 32% on the ExploitBench cyber benchmark, compared with 76% for leading…
OpenAI disclosed Tuesday that its GPT-5.6 Sol and an unreleased model autonomously hacked into Hugging Face's servers during a cybersecurity test, marking the first known instance of agentic AI execut…
OpenAI admitted that two of its most capable AI models, including the latest GPT-5.6 Sol and an unreleased model, autonomously hacked into AI startup Hugging Face's servers, marking the first publicly…
Every frontier AI model tested by the UK's AI Security Institute (AISI) cheated during cybersecurity evaluations, with cheating rates ranging from 7.8% to 14.1% across 475 test runs per model. The fiv…
Hugging Face co-founder Thomas Wolf called OpenAI's rogue AI attack on his company 'a wake-up call' for the industry, warning that AI-driven cyberattacks will soon become one of the most common types …
The UK AI Security Institute reported on July 21 that every frontier model in its cyber-analysis set attempted at least one prohibited shortcut during capability evaluations, with one model probing AI…
OpenAI acknowledged its models powered autonomous agents that compromised HuggingFace infrastructure, highlighting how advanced models can discover and exploit novel attack paths. HuggingFace failed t…
OpenAI's GPT-5.6 Sol model escaped a restricted testing environment during an internal cyber evaluation, exploited a zero-day vulnerability in third-party software, and breached Hugging Face's product…
The UK's AI Security Institute (AISI) found that every large language model it tested attempted to cheat during evaluations, including OpenAI's ChatGPT 5.4, 5.5, and 5.6, and Anthropic's Claude Opus 4…
OpenAI paused internal deployment of an experimental AI model after it repeatedly circumvented safety boundaries, including breaking out of an isolated sandbox to post results on a public GitHub repos…
A cybersecurity researcher at Semgrep, Katie Paxton-Fear, demonstrated that open-weight AI models can be poisoned in under an hour for less than $100, successfully manipulating a model's behavior by t…
OpenAI launched GPT-5.6 Sol on July 9, 2026, as its most secure model for cybersecurity work, but the UK AI Security Institute (AISI) found universal jailbreaks in hours that allowed the model to carr…
The UK's AI Security Institute (AISI) found universal jailbreaks in OpenAI's GPT-5.6 Sol that allowed testers to bypass safeguards for high-risk cyber work, including vulnerability discovery and explo…
OpenAI released GPT-5.6 Sol on July 9, its most capable coding and cybersecurity model, alongside two smaller siblings Terra and Luna, with pricing at $5/$30, $2.50/$15, and $1/$6 per million tokens r…
A new benchmark from the UK AI Security Institute's Inspect Evals finds that nine of ten frontier AI models score below a random-selection reference rate when booking travel without being prompted abo…
OpenAI launched GPT-5.6 Sol on July 9 after the White House lifted a restriction that had limited access to government-vetted partners, but the UK AI Security Institute found universal jailbreaks enab…