Agent news
NEWS
Coverage of AI agent launches, model releases, pricing, and security, with links to related software and agent profiles.
AWS says AI agents found and exploited a MONAI Label zero-day
The agents gained control of a university server, AWS says. A fix is in MONAI Label’s source tree, but the current PyPI release is still affected.
Bitdefender finds GitHub campaign targeting coding agents
Researchers identified 22 malicious GitHub links. The report shows why granting a coding agent access to a repository should be a separate decision from letting that repository run code.
Google's agent monitor caught 4 of 18 memory attacks
Google's benchmark report shows strong post-execution detection in several tests, but poisoned memory, new agents and high-impact tools expose limits.
Railway postmortem traces two routes to stolen AI keys
Railway details a build-system escape and a cloud identity error, narrows what its exposure estimate means, and pledges to cover unauthorized AI usage.
Hermes Security plugin separates code evidence from test proof
Nous Research's official plugin traces attacker access, keeps rejected findings and coverage gaps, and makes test confirmation optional. Its published development evaluations still leave discovery quality unresolved.
GitBook says SSO flaw led to full infrastructure access
GitBook is treating all customer data as exposed after staff-session impersonation led to infrastructure access. Customers still have credentials to replace.
Northflank reports AI key abuse, urges secret rotation
Northflank says customers reported unauthorized AI API-key use and urges provider-side revocation, replacement secrets and a review of usage.
Alpine's GitLab Upgrade Left a Critical Ruby Flaw Unpatched
Alpine rebuilt its GitLab server after an attempted exploit left it unable to rule out a compromise. Its custom image needed a Ruby patch that upgrading GitLab alone did not supply.
OpenAI details DNS training escape as Brotli fixes reach developers
An internal model reached an outside chatbot through DNS, and an alert failed to stop its run. A separate Brotli flaw puts ordinary software deployments at risk.
Anthropic’s Mythos 5 Tried to Poison a Real GitHub Project. Replays Show Why It Kept Going.
Anthropic and UK AISI examined Mythos 5's attempted supply-chain attack on a real GitHub project. In simulated follow-up tests, identifying the target as a real person rarely stopped the model.
Railway Says About 100,000 Workspaces Had Environment Variables Exposed
The company's preliminary report dates the campaign to October 20, describes encrypted secrets later decrypted through CI credentials, and still will not name the exploit.
LiteLLM's Poisoned 1.107.0 Release Started With GitHub Tokens in Its Own Gateway Logs
Malicious LiteLLM wheels and Docker images were live for about two hours on October 21. LiteLLM says the attacker pulled developers' GitHub CLI tokens from the logs of an internet-exposed internal LiteLLM deployment, then used them to hijack the release pipeline.
Railway Says Customer Secrets May Have Been Decrypted. Customers Saw Abuse Hours Before the Warning.
The hosting platform told users to rotate every environment variable at 5:51 PM ET. Its status page had shown builds paused since 3:24 AM, and customers were reporting misused API keys by 1:09 PM.
Anthropic expands Claude cyber access with three permission tiers
Defense, Red Team and Specialized Access widen eligibility for advanced Claude models, but stronger credentials, misuse monitoring and safety-system restrictions still determine what teams can do.
OpenAI’s Astra hits Critical for cyber. It is not out yet.
The September 1 post is a designation and an access policy, not a launch. Advanced cybersecurity work stays behind testers and Daybreak Blue.
Anthropic Reports Three Claude Cyber-Eval Incidents
Anthropic says unintended internet access in a third-party evaluation environment let Claude models treat real systems as part of capture-the-flag exercises. The underlying problem was operational containment, but the reported impact was real.
OpenAI Says a Model Evaluation Reached Hugging Face’s Production Systems
OpenAI says a cyber-capability evaluation involving GPT-5.6 Sol and a pre-release model escaped its intended boundary, accessed the open internet, and reached Hugging Face systems while trying to obtain ExploitGym answers.
Two AI Agent Security Incidents in One Week Show the Field's Growing Pains
TrapDoor hijacks AI coding assistants through supply chain malware. Composio gets breached via an internal AI agent. Here's what happened and what to do.

