DeepKeep launches AI agent attack surface scanner to map enterprise risk. DeepKeep releases an agent attack-surface scanner that maps LLM-agent risks across enterprise workflows, surfacing vulnerabilities and exposures for remediation. Outcome engineers should treat attack-surface mapping as a first-class engineering artifact — it’s how you discover where to apply guards and automated remediation (Principle 14).

Zenity warns of inherent security risks in agentic browsers after Perplexity Comet findings. Zenity discloses critical vulnerabilities in agentic browsers enabling zero-click hijacking, local file exfiltration, and credential theft. If your agents browse or act on the web, assume the browser is a high-risk surface and design strict sandboxes, credential isolation, and least-privilege flows immediately (Principle 07 / 14).

JetStream Security raises $34M seed for AI Blueprints real-time agent-mapping tool. JetStream’s AI Blueprints provides real-time mapping of agent activity to make behavior transparent for governance and audit. This is the observability layer outcome engineers need to tie agent actions back to context, policies, and artifacts so you can enforce accountability and automate post‑hoc validation (Principles 06 and 13).

OpenAI’s AI data agent, built by two engineers, now serves 4,000 employees — and the company says anyone can replicate it. OpenAI deploys a GPT-5.2-powered internal data agent giving plain‑English access to 600PB of corporate data and fast analyses. Treat this as a practical pattern: small teams can ship high-leverage data agents, but you must pair them with access controls, provenance tracking, and context engineering to avoid trust and security failures (Principles 03 and 06).

Hallmark — detect LLM hallucinations locally in Elixir. Hallmark runs a 184M entailment model locally to score LLM output grounding and catch hallucinations without sending data offsite. Local, lightweight verification layers like this are a pragmatic way to reduce hallucinations and data exposure while preserving latency — integrate them as part of your outcome-validation pipeline (Principles 02 and 16).