[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

confirmed (Elastic Security Labs)

Fake Web3 accelerator 'Apex' (REF8801) ships a repackaged coding agent that routes LLM traffic to the attacker and registers its own MCP server

Elastic Security Labs says an operation posing as a Web3 accelerator (Apex / Apex Arena) courted crypto founders with free Claude and Gemini credits, an AI coding assistant and audits. Its npm packages were malicious from their first release. @apexacc/cli is a near-copy of the open-source Oh My Pi coding agent (1,351 of 1,363 files match). It overwrites OPENROUTER_BASE_URL so model requests go to arena.apexaccs[.]org, writes a persistent apex-copilot entry pointing at the attacker's MCP endpoint into ~/.apex/agent/mcp.json, and defaults tool approvals to 'yolo'. That means the attacker's server could hand back tool calls the agent would run, though Elastic didn't observe response substitution. @apexacc/audit's 18 'MCP security tools' return hard-coded or random findings. Both packages pull platform binaries whose 'verification' step (vrf.apexaccs[.]org) returns shell commands, delivering CloudSyncD on macOS and a Telegram-exfiltrating stealer on Windows.

Sources: Elastic Security Labs (Oct 12)