Gemini API File Search is now multimodal: build efficient, verifiable RAG. Google adds multimodal retrieval, custom metadata, and page‑level citations to the Gemini file‑search API to support richer, verifiable RAG pipelines. Outcome engineers can use these primitives to build retrieval layers that carry provenance and evidence into agent decisions, reducing hallucination risk and improving auditability.

LLM Agents Find Kernel, Docker, OpenSSL Vulnerabilities. Chains of LLM agents autonomously discover remote out‑of‑bounds write bugs across the Linux kernel, Docker, and OpenSSL. That both accelerates vulnerability research and expands your threat model—agent orchestration can be a tool for discovery and a vector for misuse, so bake detection, rate limits, and governance into agent fleets.

Anthropic Limits Access to Claude Mythos Model. Anthropic restricts Mythos access and launches Project Glasswing to distribute vulnerability‑scanning to vetted partners with credits and security donations. For outcome engineers this signals tighter access controls and vetting for powerful tooling—expect gated APIs, partner programs, and new operational steps for safe testing and disclosure coordination.

Ollama contains critical GGUF out-of-bounds read. A critical GGUF model‑loader bug in Ollama (CVE‑2026‑7482) permits heap data leakage and artifact exfiltration via unauthenticated endpoints. If you run self‑hosted inference, treat model loaders and artifact formats as first‑class attack surfaces: harden access controls, validate artifacts, and include loader fuzzing in your CI.

Alibaba Integrates Qwen AI With Taobao For Agentic Shopping. Alibaba embeds Qwen into Taobao/Tmall to enable agent‑driven end‑to‑end shopping, including payments and post‑sale workflows. This is a concrete agentic commerce stack—outcome engineers building transactional agents must design for identity, atomic payments, and recovery semantics across external rails.