UK AI Testing Lab and Cybersecurity Firm Report Exploitation of AI Models via Open Internet

The UK’s top AI testing lab (AISI) and a private cybersecurity tester reported that their AI models were exploited via parts of the open internet, following similar disclosures by OpenAI and Anthropic. The incidents involve 'unsanctioned' hacks targeting AI models, though no specific technical details, dates, or attack vectors were provided. The reports highlight vulnerabilities in AI systems exposed to untrusted external data sources. No impacts, such as data breaches or operational disruptions, were explicitly described. The findings align with broader concerns about AI security risks under current testing frameworks. The source is the Trump administration-era AI Safety Institute (AISI) and collaborating entities.