Log in Sign up
Opinion Bill Gates Says AI Cyberattacks Are the Threat to Watch Next Year - and Rejects Industry Self-Regulation AI Agents Agent Teams Cost Up to 5.1x More and Almost Never Win: Vals AI Puts a Number on the Coordination Tax Business An FTX-Era Bitcoin Miner Becomes an AI Data Center Bet: SWI Group Raises 800 Million Euros at a 6.1 Billion Valuation Security Britain Moves to Force Frontier AI Developers to Report Security Breaches - and Regains Pre-Release Access to US Models Coding Assistants OpenAI Raises GPT-6.1 Sol's Default Context in Codex and Work to 500K Tokens — and Drops the Long-Context Fee for Subscribers Claude Anthropic Pulled Its Free Startup Perks 48 Hours After Launching Them — Applications Came in 21 Times Faster Than the Previous Five Months AI Agents A 1.7B Model Just Beat Its 6B and 8B Rivals: DexForce's UNITAS Treats Robot Manipulation as 3D Points, Not Pixels Business ASML Is Raising Prices 10% on Every EUV and DUV Spare Part Sold in South Korea — and Samsung and SK Hynix Have Already Agreed Opinion Bill Gates Says AI Cyberattacks Are the Threat to Watch Next Year - and Rejects Industry Self-Regulation AI Agents Agent Teams Cost Up to 5.1x More and Almost Never Win: Vals AI Puts a Number on the Coordination Tax Business An FTX-Era Bitcoin Miner Becomes an AI Data Center Bet: SWI Group Raises 800 Million Euros at a 6.1 Billion Valuation Security Britain Moves to Force Frontier AI Developers to Report Security Breaches - and Regains Pre-Release Access to US Models Coding Assistants OpenAI Raises GPT-6.1 Sol's Default Context in Codex and Work to 500K Tokens — and Drops the Long-Context Fee for Subscribers Claude Anthropic Pulled Its Free Startup Perks 48 Hours After Launching Them — Applications Came in 21 Times Faster Than the Previous Five Months AI Agents A 1.7B Model Just Beat Its 6B and 8B Rivals: DexForce's UNITAS Treats Robot Manipulation as 3D Points, Not Pixels Business ASML Is Raising Prices 10% on Every EUV and DUV Spare Part Sold in South Korea — and Samsung and SK Hynix Have Already Agreed

Britain Moves to Force Frontier AI Developers to Report Security Breaches - and Regains Pre-Release Access to US Models

Security · By: AIReport Admin · 1 hour ago · 👁 5
Britain Moves to Force Frontier AI Developers to Report Security Breaches - and Regains Pre-Release Access to US Models

UK AI minister Kanishka Narayan will announce compulsory breach reporting for frontier AI work and a national cyber defence overhaul on Monday. After lobbying from London, the AISI has regained pre-release access to models from Anthropic, Google DeepMind and OpenAI. A recent government exercise using commercial AI against its own systems surfaced roughly 1,000 vulnerabilities.

Britain is preparing to make safety reporting a legal obligation for companies building the most powerful AI systems. AI minister Kanishka Narayan is set to use a speech in west London on Monday to announce a package that pairs compulsory breach disclosure with a wider overhaul of the country's cyber defences, according to reporting by The Telegraph and The Guardian.

The centrepiece is an amendment to the Cyber Security and Resilience Bill that would require firms working on national security and critical infrastructure projects to report AI-related security breaches to Whitehall. Ministers also plan to expand the government's own use of frontier models to defend departments, and to give cybersecurity officials at the National Cyber Security Centre, the Cabinet Office and the Department for Culture, Media and Sport more resources to investigate and respond to incidents.

The speech follows a diplomatic reversal that has drawn less attention. Last month the White House asked Anthropic, OpenAI and other US developers to stop sharing unreleased models with the UK AI Security Institute, and Anthropic restricted British officials from reviewing its Claude Mythos 5.1 model. After lobbying from Narayan, that access has now been restored: AISI will again receive pre-release access to frontier models from Anthropic, Google DeepMind and OpenAI so it can evaluate risks before public deployment.

AISI is what makes the new reporting rules credible rather than symbolic. According to figures published by Parliament in April, the institute has tested more than 30 models with OpenAI, Google DeepMind and Anthropic, and its findings have prompted concrete changes before release. Its first Frontier AI Trends Report, published in December 2025, found that model success rates on apprentice-level cyber tasks had risen from under 9 percent in 2023 to about 50 percent in 2025, and that 2025 saw the first model complete an expert-level cyber task equivalent to ten years of professional experience.

The urgency is not abstract. A recent UK government exercise that used commercial AI tools to probe its own IT estate uncovered around 1,000 cybersecurity vulnerabilities. The heads of MI5, MI6 and GCHQ have separately briefed the cabinet on the risk that hostile actors could use automated attacks to shut down sensitive sites or extract confidential material. Narayan's own framing is blunt: Britain cannot rely on technology companies to police their own models, and it needs to build more of its own AI infrastructure, backed by the 500 million pound Sovereign AI venture fund whose first equity investment went to infrastructure company Callosum.

The announcement also lands in a divided international landscape. Prime Minister Andy Burnham has said Britain will use its G20 presidency to press for a single set of global AI safety principles and standards, while Washington has taken the opposite view that voluntary commitments from developers are sufficient. With AISI's access restored and reporting duties moving toward statute, Britain is betting that independent evaluation - not developer assurances - is what makes frontier AI governable.

Comments (0)

Log in to join the discussion

Log In

No comments yet