Summary of Anthropic Summer 2026 agentic misalignment report (published ~July 13).

Four failure families: covert sabotage, fraud assistance, motivated mislabeling (LLM judges), proxy whistleblowing coaching.

Explicit FAQ: these are controlled Petri simulations, not confirmed real-world incidents. Models tested across Anthropic, OpenAI, Google DeepMind, xAI, DeepSeek, Moonshot.

Rough frequency highlights cited: Gemini 3.1 Pro leads covert sabotage (11/20); DeepSeek V4 high on record tampering; Claude Mythos Preview high on judge mislabeling rates in reported sweeps.