Google ADK 2.14 graduates A2A, auth and eval APIs to stable
Pin to them without fear of breakage — plus agents caught raiding cloud secrets, OpenAI's image API gains model tiers, and Ai2 open-sources a robot stack.

Copy markdown
ADK 2.14 freezes the stable surface
Google's Agent Development Kit moves auth, environments, the A2A protocol, the core runtime and the eval classes out of experimental, so you can pin to them instead of chasing breaking changes each release. New this version: TeamsRunner (agents inside Microsoft Teams), MultiRunConfig (run eval cases concurrently) and ToolArgsMatcher (exact, numeric-tolerance or LLM-semantic argument checks). Google clocks long multi-turn workflows 23% faster with 9% lower peak memory, across Python, TypeScript, Go, Java and Kotlin SDKs.
Agents are the new secrets-exfil tool
Microsoft Threat Intelligence reports attackers pointing agentic AI at public GitHub repos to find exposed credentials and pivot straight into cloud accounts — one intrusion went from spotting a secret to pulling data out of Azure Key Vault in about five minutes. If your agents or CI touch real tokens, rotate anything that's been exposed, move secrets into a manager, and scope agent tool permissions to least-privilege now.
Update: OpenAI's image API gets tiers and a price sheet
GPT-Image-2.6 now splits into two callable models — gpt-image-2.6-flare (fast, roughly six seconds at low quality, built for iteration) and gpt-image-2.6-sunburst (consistency across repeated edits) — plus a complexity dial (low/medium/high, defaults to high, no surcharge) that trades planning time for cost. Image tokens run $8 in / $30 out per 1M, about $0.025 per 1024-square render at low complexity versus $0.053 at high. Reach for flare at low on everyday edits and save the headroom for finals.
Ai2 open-sources a train-in-sim robot stack
Ai2's ToyBlocks ships an open simulation framework with roughly 85,000 auto-generated manipulation demonstrations plus training code and Apache-licensed weights, built to train tabletop robot policies entirely in simulation. Reported atomic-skill success lands around 81%, with the usual generalization drop when block positions move. If you've wanted to prototype a manipulation policy without a real-robot data pipeline, this is a weekend starting point.