Researchers disclosed EchoLeak (CVE-2025-32711), the first documented zero-click prompt-injection exploit against a production LLM assistant. A single crafted email could make Microsoft 365 Copilot pull internal content across its trust boundaries and send it to an attacker-controlled endpoint — no click, no attachment, no credentials. Microsoft's advisory scores it 9.3 CRITICAL with a vector of no privileges and no user interaction required, and NVD lists it as an exclusively hosted service, meaning it was remediated server-side rather than patched on customer machines.

The chain bypassed Microsoft's XPIA prompt-injection classifier and then abused a content-security-policy gap via a markdown image reference to an allowed domain — ordinary web-app attack classes reappearing inside the model's trust boundary. The follow-up paper recommends prompt partitioning, provenance-based access control, strict CSP and input and output filtering.

Mistral released Magistral, its first reasoning family: Magistral Small, a 24B open-weight model licensed Apache 2.0, alongside the hosted Magistral Medium. Small builds on Mistral Small 3.1, exposes a multilingual chain-of-thought you can read, and carries a 128k context window; the model card recommends capping it at 40k, where quality starts to degrade. On Mistral's own numbers it scores 70.7% AIME24 pass@1 and 68.2% GPQA Diamond, against 73.6% and 70.8% for the closed Medium.

Mistral ships it to run locally on a single RTX 4090 or a 32GB MacBook once quantized, which is the number that matters: reasoning tokens with no per-token bill, no rate limit and no data leaving your machine. The accompanying paper publishes the RL recipe and training infrastructure next to the weights, not just the benchmarks, under permissive terms that allow commercial use and modification.

OpenAI dropped o3's API price from $10/$40 to $2/$8 per million input and output tokens — an 80% cut achieved by optimizing the inference stack, with the model itself unchanged — and launched o3-pro at $20/$80 per million, which OpenAI notes is 87% cheaper than the o1-pro it replaces. o3-pro went live to ChatGPT Pro and Team users and in the API the same afternoon.

The pro tier buys reliability with test-time compute, not new capabilities: no image generation, no Canvas, and OpenAI recommends background mode because long runs will otherwise time out. OpenAI's own evaluations put it ahead of Gemini 2.5 Pro on AIME 2024 and of Claude Opus 4 on GPQA Diamond — vendor-reported numbers, worth treating as directional.

At WWDC25 Apple gave developers direct access to the on-device foundation model behind Apple Intelligence through the Foundation Models framework: a compact roughly 3B-parameter model that runs locally, works offline, and in Apple's words uses AI inference that is free of cost. It is a Swift API with guided generation and tool calling, shipped in the same release cycle as iOS 26 and macOS 26.

Apple also detailed the model updates: a re-architected on-device model that cuts KV-cache memory 37.5% for faster time-to-first-token, better tool use and a 15-language target, plus a parallel-track mixture-of-experts server model running inside Private Cloud Compute. In the same week Apple's Illusion of Thinking research drew coverage for finding that reasoning models can collapse to near-zero accuracy on puzzles past a complexity threshold.