Cybersecurity, AI, etc.
Artificial intelligence security (LLM/agentic system vulnerabilities, AI-enabled attacks, model security research, prompt injection, data poisoning)
What happened: The Indian Express reported that Z.ai disabled features in its AI coding assistant following a security issue. The report underscores the sensitivity of coding assistants, which may handle…
Researchers chained code execution in OpenAI’s community forum with an over-privileged OpenAI sign-in token, reaching an employee account connected to internal repositories before the flaws were fixed.
A Gemini model unintentionally accessed systems belonging to three real companies after an AI security evaluation was given internet access and confused real organizations with a fictional test target.
Researchers disclosed two patched OpenAI Codex sandbox escapes, including a flaw that could let a malicious repository trigger commands on a developer’s host from the strictest sandbox mode.
Mandiant describes a compromised AI coding session that led to a supply-chain worm spreading across repositories.
OpenAI introduced a framework and six reports covering observed unexpected or concerning model behavior.
Researchers attribute a RubyGems package campaign to OpenAI agents and report RCE on RubyDoc build infrastructure.
Anthropic reports disrupting AI misuse across cyber operations, surveillance, fraud, influence, weapons, and model distillation.
GreyNoise reports an AI-orchestrated PaperCut campaign compromised at least 440 instances across 395 organizations.
Wiz found 9.6% of sampled public LiteLLM instances accepted default or no authentication, enabling severe AI-gateway compromise paths.