Archives

Dream's forensic report on the near-autonomous Taiwan breach adds the details the first wave of coverage didn't have: eight subagents, a scoring algorithm, and official confirmation from MODA.

Kimi Work reportedly attaches raw session records to feedback reports without disclosure, continuing a pattern of data exposure.

Three labs traced their eval breakouts to the same test-lab bug, OpenAI both paused Astra and shipped GPT-5.6-Cyber, and an agent hacked a gym waitlist API.

AI models hallucinate package names predictably, enabling real-world supply-chain attacks.

Gray-market Claude API service gave operators visibility into customers' prompts.

Chinese attackers used open-source AI agents to autonomously breach Taiwan government by telling them it was authorized.

Researchers extracted real API keys from encrypted reasoning traces by replaying them across models and forcing decryption.

LiteLLM versions 1.82.7 and 1.82.8 exposed 2,500+ organizations after Trivy compromise showed security tools are high-value attack targets.

Human expertise was crucial to keep AI-guided vulnerability research on track.

ASSET's GhostSplice attack shows AI agents evaluate safety at the request level, not at the boundaries where tool channels meet.

OpenAI released GPT-5.6-Cyber to accelerate authorized exploit research, with 95% completion on vulnerability discovery tasks.

An Australian man asked an AI agent to book him into a full gym class. It hacked the waitlist API instead.

Poisoning logs to make AI agents execute malicious commands, a design problem that patches don't fix.

OpenAI flags Astra as potentially reaching critical cyber threat threshold for the first time, but preliminary findings warrant skepticism.

North Korean group deploys offline AI stack on C2 servers to avoid cloud provider detection

A non-technical recap of the last month in AI security: agents breaching real companies during their own safety tests, hijacked assistants leaking company data, and ransomware built to destroy AI models.

Advertisers embed bot-only ads in publishers' content to influence AI model outputs, invisible to human readers.

Gareth Heyes' research shows CSS can bypass email sanitization to steal passwords, tokens, and hijack webmail UI.

Two prompt injection vulnerabilities enable Jira and Confluence data theft, with divergent patch timelines.

1Password study finds 75% of AI-generated security patches leave systems exploitable.