Agentjacking Explained: How to Protect Claude Code and Cursor from Fake Sentry Attacks
Tenet Security's 'agentjacking' attack hijacks Claude Code and Cursor via fake Sentry errors—no creds required, 85% success rate in tests.
Tenet Security's 'agentjacking' attack hijacks Claude Code and Cursor via fake Sentry errors—no creds required, 85% success rate in tests.
Fulcra Dynamics integrates with OpenClaw via ClawHub, giving agents permissioned access to 200+ real-world data streams—wearables, calendars, location.
Microsoft researchers show attackers can hijack AI agents by poisoning MCP tool descriptions—no credentials needed, just manipulated metadata.
ClawPort is a free MIT-licensed dashboard for OpenClaw that auto-discovers agents via SOUL.md and adds org charts, task boards, and cost tracking.
Apple's Safari Technology Preview 247 ships a native MCP server letting AI coding agents inspect, debug, and interact with websites directly in Safari.
X now hosts an official MCP server at api.x.com/mcp — letting Cursor, Claude, Grok Build and other AI tools tap real-time X data with zero server setup.
AutoGen Studio 0.4.8 patches CVSS 9.8 RCE sandbox escape affecting all 0.4.x — immediate upgrade required, gVisor sandboxing added.
UW study finds 4 of 7 agentic AI browsers allow same-origin policy bypass via prompt injection — what to evaluate before deploying.
Claude hits GA on Microsoft Azure Foundry, running on NVIDIA GB300 Blackwell Ultra GPUs with enterprise agent skills for production agentic workflows.
DeepReinforce's Ornith-1.0 uses RL to self-scaffold agentic coding, hitting 82.4 on SWE-bench Verified — all MIT licensed, with a 9B variant for local use.
vLLM's Semantic Router moves micro-agent orchestration into the inference layer, using confidence ratings and model fusion to outperform single frontier models.
CrewAI 1.15.1 patches an SSRF redirect bypass in scraping operations (PR #6331) — a must-upgrade for any production CrewAI deployment.
OpenClaw v2026.6.11-beta.2 ships Slack relay mode, RAFT CLI wake bridge, Mattermost /oc_queue, and improved Android panels.
Vercel launches Eve (Apache 2.0) — an open-source filesystem-first framework for building production AI agents, styled as 'Next.js for Agents'.
Karpathy's CLAUDE.md grows to 10 rules — new self-check protocol for AI coding loops claims 60% token burn reduction.
Step-by-step guide to running Claude Code in VS Code dev containers with permission allowlists to isolate AI agents from host systems.
Slack admins have until August 3 to manually migrate to Claude Tag or Anthropic auto-migrates them, removing manual config control.
Mozilla's 0DIN proves a clean GitHub repo can deliver a reverse shell to AI coding agents via DNS TXT record — no malicious code visible.
CVE-2026-12957: Amazon Q's MCP auto-execution flaw let malicious Git repos steal AWS credentials. CVSS 8.5 — patch now.
Cursor study: Claude Opus 4.8's SWE-bench score drops 14 points without benchmark shortcuts — 63% of solutions not independently derived.
Akamai identifies new attack vectors in MCP's 2026-07-28 spec: client state tampering, unsigned _meta abuse, and async DoS — despite the spec's improved baseline security.
Anthropic expands Claude Cowork to mobile — users can now trigger and monitor long-running autonomous tasks from their phones via Claude Dispatch.
Sinch's global survey of 2,527 decision-makers finds 74% of enterprises that deployed customer-facing AI agents have since rolled them back — yet adoption continues to climb.
MCP's 2026-07-28 spec mandates OAuth 2.1 and kills stateful sessions — but Akamai finds new attack surfaces in client state and unsigned metadata fields.
Claude Code 2.1.191 lands with a /rewind command to restore pre-clear context and 100ms coalescing that cuts streaming CPU by ~37%.