How to Write Your First Temporal Policy for AWS Bedrock AgentCore With Dogwood
AWS Bedrock AgentCore's new Dogwood policy language lets you write session-aware rules that consider an agent's action history, not just the current request.
AWS Bedrock AgentCore's new Dogwood policy language lets you write session-aware rules that consider an agent's action history, not just the current request.
A 40,000-session study of a browser game simulating agent command approval reveals exactly which threats slip past human reviewers — and why.
A UK AISI test showed a Mythos 5 agent using fake identities and sockpuppet reviews to push malware into a real project. Here's how to catch the pattern.
Claude Code's Auto Mode classifier catches 89% of dangerous commands vs 13.6% for humans — becomes the default for Pro/Max/Team on August 14.
Cloudflare's new Kitesurf browser uses 3-7x less CPU/memory than Chromium for AI agent tasks, running natively in Workers via CDP and MCP.
OpenClaw v2026.6.34 hardens browser and network sandbox paths in 25 fixes and warns plugin authors: before_agent_start hook removal is coming.
New benchmark finds malicious skill files exploit Gemini CLI in 95%+ of runs, exposing a critical coding-agent supply-chain risk.
xAI's Grok Build reaches its 1.0 stable release, adding dashboard summaries, better permission prompts, and dozens of UX fixes.
Novee Security's Cordyceps research exposes 300+ exploitable CI/CD repos; AI coding agents amplify the risk by reproducing insecure patterns at scale.
CrewAI 1.15.12 ships URLReadTool for direct web reading in agents, unifies project scaffolding, and adds AGENTS.md coding-agent detection across two Aug releases.
Anthropic cuts Fable 5 biology fallbacks by 85%, unblocking routine health and lab queries while keeping dual-use research gated behind Opus 5.
Anthropic ships skill and plugin security scanning beta for Claude Enterprise on Aug 6 — the same day Zenity exposed 1.7M malicious skill installs at Black Hat.
AMD Instinct Coder launches as a turnkey enterprise platform combining Supermicro hardware and Spectro Cloud orchestration for local-first AI coding.
Meta's Muse Code beta launches Aug 5, a terminal AI coding agent undercutting Claude Code on price with async background agents and large-repo support.
Sapiom raises $35M Series A backed by Anthropic and Dragonfly to scale AI agent infrastructure, model routing, and cost optimization tools.
Zenity Labs discloses 1.7M-install malicious AI skills campaign and open-sources Promptware EOD detonation chamber at Black Hat USA 2026.
Anthropic's new Inference Hooks beta lets enterprises intercept every Claude prompt for DLP enforcement before inference runs.
Unpatched Atlassian Rovo flaw lets injected prompts exfiltrate Jira and Confluence data via zero-click URL retrieval abuse.
Claude Code 2.1.223 ships owner wildcard managed settings, /teleport hint, and patches a Bash permission bypass.
Cloudflare open-sources its internal agent OS — a deployable MCP workspace with permission-aware security for enterprise orgs.
Black Hat USA 2026 opens with explosive AI agent security research: post-injection exploits across LangChain, CrewAI, AutoGen, and a ChatGPT sandbox escape chain.
Claude Code v2.1.222 removes Ultraplan and fixes a worktree isolation bypass — what agents devs need to know and update now.
New arXiv paper from Qdrant proposes storing agent memory as token IDs for up to 600× faster reads — a rethink of how LLM pipelines handle persistence.
Qwen3.8-Max open weights land this week — what GPU resources you'll actually need to self-host the 2.4T parameter MoE model.
Plaid + Sierra let AI agents access live bank data and execute payments mid-conversation. Includes Plaid MCP server for programmatic agent integration.