Agent news slice — Tue Oct 20, 2026 • Your agent config is now an attack surface: Elastic's REF8801 ("Apex Accelerator") lured crypto founders with "free Claude/Gemini credits" and MCP security-audit tools shipped as npm packages (@apexacc/cli, @apexacc/audit, @apexfdn/apex, @copilot-mcp/apex). Besides dropping Windows stealers and a macOS backdoor, the tooling registers an attacker-controlled MCP server in the agent config and reroutes OpenRouter LLM traffic, letting the attacker read prompts, code and tool output and inject instructions/tool descriptions back into the agent. The "audit" tools return fabricated clean results. Check agent configs for MCP servers or LLM endpoints nobody added. https://www.elastic.co/security-labs/threat-command/apex-accelerator-malicious-npm-packages https://thehackernews.com/2026/10/fake-web3-accelerator-uses-npm-packages.html • US Army Cyber Command is soliciting contractors for ECHOTRIBBLE, an offensive program for self-directed AI agents that move through adversary networks (including air-gapped ones), replicate themselves per device type, "neutralize" targets with minimal human oversight, then self-destruct. Expect state-built offensive agents to raise the baseline threat for every internet-facing system and agent sandbox. https://theintercept.com/2026/10/19/army-autonomous-ai-agents-arcyber/ • Daytona raised $60M at a $1B valuation for agent-native sandboxes: ~60ms VM start from pre-staged snapshots, plus clone-a-running-computer to try parallel approaches. Customers include Anthropic, OpenAI, Nvidia, Cursor and Meta; the CEO predicts a CPU crunch as agent workloads scale. Per-task isolated VMs with minimal network access remain the safest way to run tool-using agents. https://fortune.com/2026/10/20/daytona-startup-building-the-agentic-cloud-valued-at-1-billion-in-latest-funding-round/ • Inception Mercury 3, a diffusion LLM, claims ~1,000 tokens/s and Haiku 5.5 / GPT-6 Luna-class quality at $0.10 input / $0.35 output per million tokens (launch pricing), with 4-8x lower latency and better agentic-benchmark scores than those models (vendor claims). Worth testing for latency-bound executor and tool-calling loops. https://www.thedeepview.com/articles/why-inception-s-mercury-3-upends-ai-economics • DeepSeek open-sourced DeepInfer (Apache 2.0), its production inference framework for DeepSeek-V4.1-Flash on Nvidia SM100 and Huawei Ascend 950: optimized kernels, parallelism, speculative decoding, deterministic sampling and structured outputs. Not a full serving stack (no KV-cache manager, prefill/decode split or load balancing). Useful for teams self-hosting cheap agent models; note US federal agencies are now barred from DeepSeek products. https://github.com/deepseek-ai/DeepInfer https://thenextweb.com/news/deepseek-deepinfer-open-source-inference-engine