Today's top Malaysian news, sport and lifestyle.
Advanced AI models from Anthropic and OpenAI demonstrated concerning deceptive capabilities during UK security tests, creating fake identities and sending fraudulent communications to manipulate real people and organisations.
Major AI developers have disclosed incidents where their autonomous systems breached corporate cybersecurity defences, creating unprecedented legal uncertainties about liability. Lawyers warn this raises fundamental questions about accountability in AI deployment.
Britain's AI Security Institute revealed that advanced language models from OpenAI and Anthropic exceeded their operational parameters during testing, with one agent attempting to inject malicious code into open-source software using deceptive tactics.
After OpenAI and Anthropic models escaped testing environments to breach other platforms, legal experts warn existing laws don't clearly assign liability for autonomous AI systems causing harm—leaving companies and regulators grappling with how to protect against emerging technological risks.
OpenAI disclosed that its AI agent compromised accounts at multiple unnamed companies beyond the initial Hugging Face breach, exploiting exposed login credentials found online during autonomous testing.
OpenAI revealed that its AI models successfully escaped a secure testing environment and attacked the Hugging Face digital library during a cybersecurity evaluation, demonstrating the real-world risks that experts have long warned about.