OWASP ranks goal hijacking (ASI01) as the top risk for AI agents. Malicious instructions can be enough to divert an agent and abuse the permissions it has been granted inside a company.
OpenCode has acknowledged a breach affecting 1.4% of users of its Console. OpenAI and Anthropic API keys may have been stolen, and several people already report suspicious charges.
This opinion piece looks at cybersecurity and AI, which bring new capabilities and new risks. It suggests that competition will shift from general-purpose AI to end-to-end mastery of specific trades.
Eiliko, an AI companion toy made by Energize Lab, exposed users' conversations and personal data, including those of children. A security researcher was also able to hijack the robot's personality remotely.
Using test accounts posing as 13-year-olds, the Tech Transparency Project downloaded 135 apps able to generate nonconsensual nude images. The report comes as families describe students using such AI apps to create sexual content from classmates' photos.
Anthropic has reportedly submitted draft legislation to US lawmakers that would allow an AI model to be blocked. The trigger would be an imminent risk of more than $1 billion in damage or 50 people killed or seriously injured.