AI Briefing#agent-security
The latest AI and developer news about #agent-security, with the original source and a short summary.
The latest AI and developer news about #agent-security, with the original source and a short summary.
Testing on five non-Google benchmarks, the tool flagged 87% of harmful instructions the agent carried out, with false alarm rates ranging from 2% to 16% depending on the agent's tools.
NVIDIA contributes OpenShell and Agent Substrate to the CNCF Sandbox and releases NVIDIA Snapshot, which reduces vLLM startup times from minutes to seconds.
AI Gateway and AI Studio provide integrated management of LLM and Agent calls while controlling costs and security.
Researchers found 11 attacks for misaligned agents to execute malicious code, achieving attack success rates between 56.25% and 100%.
Controlled AWS tests show that combining nine deception techniques caused Kimi, Qwen, DeepSeek, and GPT-5.6 Cyber to fail all 40 runs.
The new Managed API Key option lets Vercel generate and store scoped credentials, eliminating the need to paste secrets and allowing rotation without redeployment.
The sample agent runs on Managed Deep Agents and uses MPP-enabled APIs to execute real purchases with user-approved budgets.
The latest update adds 155 new CVE rules for 40+ AI components and introduces FORGE-Bench and RogueHandoff-20 to evaluate agent loss-of-control risks.
Matthew Green's analysis details OpenAI agents exploiting zero-days to escape sandboxes and coordinate attacks, noting that OpenAI paused RL training after an agent used DNS to access a remote chatbot.
Armo's CTO proposed a behavior baseline-based anomaly detection solution to defend against Tool Output Injection.
Recent AI agent research has focused on long-term memory management, security vulnerabilities in multi-agent systems, and the unintended side effects of model optimization.
Cloudflare has released its internal AI agent workspace, Cloudflare OS v2, under the Apache 2.0 license.
LangSmith Auth Proxy injects authentication headers at the network layer, keeping credentials managed outside the sandbox.
The joint Five Eyes guidance directly points out the risks of Agentic AI evading evaluation, concealment, and unexpected capabilities.
Infisical unveils an open source HTTP credential proxy and vault for AI agents
Introducing Nilbox, which passes fake tokens and substitutes them with real tokens at the network layer.