AI TrackerInfrastructure news,
for builders.
The latest infrastructure news from every AI provider: launches, price changes and retirements, summarized with what they mean for you.
Microsoft brings the Copilot harness to Foundry Agent Service
Microsoft introduced the Microsoft Copilot harness in Foundry Agent Service. Microsoft says GitHub's engineering team built a multi-model harness that it has now implemented across Microsoft products, and that the Foundry version lets developers build agents without rebuilding the execution layer. According to Microsoft, the harness handles orchestration, tools, context and controls, so developers can focus on their agent's skills and outcomes. Microsoft published a demo video and a Foundry blog post.
Vercel AI Gateway adds 14 ElevenLabs models, 50% off through November 9
Vercel says 14 ElevenLabs models are now available on AI Gateway, its unified endpoint for model providers, and are 50% off through November 9. Vercel lists realtime voice agents, music, sound effects, multi-voice dialogue, streaming speech over WebSockets and speech to text among them.Vercel's HarnessAgent now accepts image input
Vercel says HarnessAgent now supports image input across every harness adapter. You pass images alongside text in the standard AI SDK message format, and according to Vercel each harness's underlying vision model handles the rest. The details are in Vercel's changelog entry.Parallel launches the Decisions API, which grounds decision models in current web data at $0.001 per request
Parallel announced the Decisions API on X on October 27 and says it is live. It calls it "the first web-grounded decisions engine": decision models such as Jev answer instantly but lack the recency of current news, and the new API combines both in a single call. That claim is Parallel's own.
Anthropic's Logan Graham corrects the Washington Post account of the Cathedral intrusion
Logan Graham of Anthropic's Frontier Red Team posted on October 26 corrections to the Washington Post story and the reaction to it on X, which said a security startup's AI agents got into Anthropic's systems in under 12 hours. These are his statements, not an Anthropic announcement:Cloudflare says a GLM 5.3 setup mismatch on Workers AI exposed KV cache pages across requests
Cloudflare published a disclosure and remediation write-up on October 26 saying a configuration mismatch in GLM 5.3 (@cf/zai-org/glm-5.3) on Workers AI triggered a bug in SGLang, the inference engine, that exposed KV cache pages across requests. The disclosure was coordinated with a post from SGLang.Scale AI's AgentEnv evals can now run on Vercel Sandbox
Vercel says you can now run AgentEnv evals from Scale AI Labs on Vercel Sandbox. You give an agent a task, run it in a sandbox and check how it performed. Vercel published a changelog entry and a setup guide in its sandbox documentation.Vercel Sandbox commands now support stdin
Vercel says commands run in Vercel Sandbox can now receive input on stdin. That lets you send input to running CLI tools and to stdio MCP servers, such as agent-browser, inside a sandbox. The details are in Vercel's changelog entry.Gemini API now stores Generate Content requests by default on the paid tier
Google's Gemini API changelog says new projects using the Generate Content API (generateContent, streamGenerateContent and batchGenerateContent) on the paid tier now store requests by default (store=true), matching the Interactions API. Existing projects keep their current setting.Washington Post reports a security startup's AI agents got into Anthropic's internal systems in under 12 hours
Elizabeth Dwoskin of The Washington Post posted on October 26 an exclusive saying that a group of former DOGE cyber researchers used AI agents to penetrate Anthropic's systems in less than 12 hours, including Slack channels where the company's most sensitive models are discussed. Her post ties the story to a recent charge by Anthropic whistleblower Joseph Coxon.Together AI opens MiniMax M3.1 Flash for post-training
Together AI says MiniMax M3.1 Flash is now available for post-training on its platform. According to Together, users can run supervised fine-tuning (SFT) or reinforcement learning (RL) with their own data, rewards and environments, while keeping the training loop in their own Python code.A seller claims to have OpenRouter's user database and API keys; OpenRouter finds no evidence
The threat-intelligence account DailyDarkWeb reported on October 25 that a seller on an underground forum claims to be offering a database tied to openrouter.ai. The account says the seller claims:Researcher publishes 11 unpatched flaws in Tailscale and headscale, mostly in Tailscale SSH
Erica Windisch of Cyberdione Labs published 11 findings under the name Tailworm (tailworm.com) without coordinated disclosure, so the vendor had no time to patch first. No patches or CVEs had been reported by October 25.
SGLang says it now matches DeepSeek's DeepInfer engine on DeepSeek-V4.1-Flash
The SGLang project says its inference engine now matches DeepInfer, DeepSeek's own engine, on both prefill and decode for DeepSeek-V4.1-Flash. A few hours after vLLM made a similar claim, SGLang reported up to 1.04× DeepInfer's prefill throughput on 4× GB300 and 4× B300, and up to 7.4% decode throughput on 64× GB300.
vLLM says it now matches DeepSeek's DeepInfer engine on DeepSeek-V4.1-Flash
The vLLM project says its inference engine now matches DeepInfer, DeepSeek's own engine for DeepSeek-V4.1-Flash, on DeepSeek's decode configuration across batch sizes 8 to 256. It says it does so at DEP32 (data-parallel attention with expert parallelism over 32 GPUs), against DeepInfer's EP64DP64 setup.