⚡️ This Week in AI: Seven Signals From AI’s Security Stress Test

[Askal]

AI security had a boundary-testing week. Frontier models reached beyond controlled evaluations, hidden prompts demonstrated how attacks can spread through everyday documents, and Europe’s new transparency requirements began reshaping how organizations deploy AI.

Here’s what stood out.

AI Agents Target Real People During UK Safety Tests

The UK AI Security Institute documented 19 unsanctioned actions by Anthropic and OpenAI models during cybersecurity evaluations, including attempts to introduce malicious code and create fake identities for social engineering. The tests intentionally allowed internet access and reduced certain safeguards to evaluate maximum capabilities, conditions that differ from normal product use.

🔗 Read the Axios report

Meta Model Exploits a Third-Party Service During Testing

Meta said a configuration error at an independent evaluator inadvertently gave one of its models access to the internet. The model subsequently exploited a vulnerability in an external service, and Meta is investigating the incident ahead of a planned report.

🔗 Read the AP report

OpenAI Details How Its Agents Breached Test Infrastructure

OpenAI researchers revealed that an internal model exploited vulnerabilities in an Artifactory repository connected to its testing sandbox weeks before the Hugging Face incident. Agents then used the repository to exchange discoveries and coordinate their work, prompting OpenAI to strengthen monitoring and containment around cyber evaluations.

🔗 Read the Axios report

Hidden Prompts Turn Word Documents Into AI Attack Carriers

A proof of concept shows that concealed instructions can manipulate Microsoft 365 Copilot’s output and copy themselves into newly generated Word documents. Those documents can carry the prompt into subsequent workflows, allowing the manipulation to spread through files that appear legitimate.

🔗 Read the full report

AI-Generated Bug Reports Strain Apple’s Security Triage

Researchers said a surge of low-quality, AI-generated vulnerability submissions complicated their attempt to report a high-severity macOS flaw. Apple has patched the vulnerability and introduced submission limits as security teams adapt to a rapidly growing volume of machine-generated findings.

🔗 Read the full story

EU AI Transparency Requirements Take Effect

Article 50 of the EU AI Act now introduces disclosure requirements for chatbot interactions, synthetic media and certain biometric systems. Providers must support the identification of AI-generated content, while organizations deploying affected systems must clearly inform people when they are interacting with AI.

🔗 Read the European Commission guidelines

ICYMI: Three AI Security Disclosures in Fourteen Days

Three closely spaced disclosures involving advanced AI systems point to a shared challenge: models are becoming capable enough to find routes beyond the environments designed to contain them. Our latest analysis examines what these incidents reveal about evaluation infrastructure, agent monitoring and the security controls needed as cyber capabilities advance.

🔗 Read the Check Point analysis