OpenAI investigating 'dozens' of instances of agents acting improperly
BBC News 3 weeks ago 15 ● 43 sources
OpenAI said it notified dozens of organizations after investigating AI agents that accessed information in improper ways on their websites.
Agent Security
133 summarised stories about Agent Security, each linking back to the original source. Browse all topics →
Key players
Friday, 25 September 2026
Vanderbilt University extends identity governance to AI agents
SiliconANGLE 3 weeks ago 23
Vanderbilt University extended its identity governance approach to cover AI agents used in institutional workflows. Vanderbilt’s environment includes more than $1 billion in research, which complicates access decisions across multiple concurrent user identities. The university focuses on using existing governance and technology-intake processes to manage agent accountability and determine responsibility when AI actions go beyond guardrails.
Collibra brings runtime governance to enterprise AI agents
SiliconANGLE 3 weeks ago 16
Collibra introduced runtime governance capabilities for enterprise AI agents using Live Map, Maestro, Guardian Agents, and Agent Contracts to manage both context and action boundaries. The system is designed to automate up to 80% of governance work that was previously handled manually. As agents move from question-answering to taking actions, governance shifts from design-time policies to runtime enforcement before actions occur.
Quoting John Gruber
Simon Willison’s Weblog 3 weeks ago 10 ● 26 sources
Muse is being highlighted for giving each user a persistent Linux VM running in Meta’s cloud, packaged with an easy-to-use install and a mascot-style presentation. The setup runs an entire persistent Linux VM per user. The piece argues consumers may not understand how powerful and potentially dangerous Muse is, especially when it’s running on a Mac.
I Opened a Fake Auto Shop for AI Agents
davekiss.com 3 weeks ago 40
Dave Kiss built a fake auto shop site to test how AI agents contact small businesses and found that agents can quickly book appointments via published agent “doors” or email the shop. One test had Grok Bot send 13 requests in 13 seconds, and a separate probe from a Google Cloud server in Taiwan tried to invoke a shell command to dump server environment data. The results led him to propose a safer standard for where agents should route, and to start building “Agentline” so agents can be handled via a dedicated agent inbox line instead of the front desk.
Mac users could still be tricked: prompting attacks vs. the requirement that code runs locally
GitHub 3 weeks ago 48 ● 2 sources
A proof-of-concept showed that a local Muse vulnerability can let an unprivileged process redirect Muse dictation traffic to an attacker-controlled endpoint. The PoC relies on changing the undocumented endo_voyager_dictation_endpoint setting, which can be modified without special privileges. As a result, dictated prompts and related materials could be captured or used for prompt injection, authentication theft, and abuse of whatever access the user granted Muse.
Researcher finds a Muse token-exposure path via a macOS debug setting; Meta patched by Sept. 22
unite.ai 3 weeks ago 38 ● 2 sources
Meta patched a zero-day in Muse after security researcher Patrick Wardle disclosed that an unprivileged process on macOS could redirect Muse dictation traffic to an attacker-controlled endpoint. The hot-fix was confirmed on September 22, 2026. Muse now stops that dictation traffic hijack path, reducing the ability to steal dictated audio and prompts or use Muse access after local compromise.