AI
AI Agents, Open Models, and Government Security Tests
From multi-agent orchestrators to local AI powerhouses, today's HeadFlash covers breakthroughs and security revelations.
Sakana Fugu Debuts Multi-Agent System as a Single Model API
Sakana AI launched Fugu, a multi-agent system delivered as a single model API that dynamically orchestrates a pool of powerful LLMs. Grounded in two ICLR 2026 papers - TRINITY and the Conductor - it uses a lightweight evolved coordinator and reinforcement learning to assign roles like Thinker, Worker, or Verifier across models without hand-designed workflows. Two versions are available: Fugu for everyday tasks and Fugu Ultra for high-stakes work, both accessible via an OpenAI-compatible API. Users can opt specific agents out of the pool for data and privacy compliance, though Fugu Ultra’s pool is fixed. Quantitative benchmarks show Fugu Ultra leading on several metrics: SWE Bench Pro 73.7, TerminalBench 2.1 at 82.1, LiveCodeBench 93.2, and Humanity’s Last Exam at 50.0, often outperforming Opus 4.8, Gemini 3.1 Pro, and GPT 5.5. Qualitative examples include autonomous research improving a GPT recipe, solving Rubik’s cubes, and a blindfold chess win against Stockfish. Pricing ranges from $20/month Standard to $200/month Max, with pay-as-you-go rates for Fugu Ultra at $5 input and $30 output per million tokens. The system is not available in the EU/EEA due to GDPR compliance work.
Sakana Fugu — Multi-Agent System as a Model →
Datalab Releases Lift: Open-Weight Vision Model for Structured PDF Extraction
Datalab released Lift, a 9B-parameter open-weights vision model designed to extract structured JSON from PDFs and images using a provided JSON Schema. The model handles multi-page documents in a single pass and uses schema-constrained decoding to guarantee valid JSON output, though correctness of field values is not assured. Two inference modes are available: local via HuggingFace and production-ready via vLLM. Code is Apache 2.0; weights use a modified OpenRAIL-M license free for research, personal use, and startups under $5 million, prohibiting competition with Datalab’s API. On Datalab’s 225-document benchmark, Lift achieved 90.2% field accuracy and 20.9% full-document accuracy with a median latency of 9.5 seconds, outperforming self-hostable models like NuExtract3 and Qwen3.5-9B on field accuracy. Full-document accuracy remains low across all models; Datalab recommends Lift for field-level extraction feeding human review or aggregate analytics, not zero-touch automation. Supported GPUs include H100, A100, L40S, and consumer cards like 4090 and 3090.
Anthropic Launches Claude Tag: Persistent AI Teammate in Slack
Anthropic introduced Claude Tag on Tuesday, embedding its Claude Opus 4.8 model directly into Slack as a persistent, shared teammate that any team member can delegate work to by typing @Claude. Available in beta for Claude Enterprise and Team customers, it replaces the existing Claude in Slack app. The system is multiplayer - one Claude interacts with everyone in a given channel, learning over time and accumulating context. With ambient behavior enabled, it proactively surfaces information and works asynchronously over hours or days. Anthropic reports that 65% of its own product team’s code is now created by its internal version of Claude Tag. Administrators can set token-spend limits and review a complete action log. Migration from the old app requires opt-in within 30 days. Claude Tag enters a competitive landscape including Salesforce’s Slackbot, OpenAI’s Workspace Agents, and Perplexity’s enterprise agent. Anthropic’s financial backdrop includes a $65 billion Series H at a $965 billion valuation and over $47 billion run-rate revenue. Risks include increased vendor dependency, governance questions around ambient monitoring, and potential infrastructure strain from surging demand. Anthropic plans to expand Claude Tag beyond Slack to other platforms like Microsoft Teams and email.
Chinese AI Model GLM 5.2 Runs Locally, Outperforms ChatGPT
Developers are adopting GLM 5.2, an open-weight AI model from Z.ai (formerly Zhipu AI) that can be downloaded, customized, and run entirely on local hardware. The model challenges the assumption that frontier performance requires premium subscriptions, as it offers direct access to model weights. GLM 5.2 is a 744-billion to 753-billion parameter Mixture-of-Experts model with a 1-million-token context window, similar to Claude. Its uncompressed weights consume 1.51 terabytes of storage and memory. Even when heavily quantized, it requires roughly 240GB of memory to load, putting it beyond standard high-end PCs but within reach of maxed Mac Studio configurations (256GB unified memory). Developers have demonstrated it running on high-end Apple hardware including the Mac mini. For industries handling sensitive data, local operation offers a security advantage by keeping data on private hardware.
This new Chinese AI Is outperforming ChatGPT — and it runs locally →
Anthropic Mythos Finds Vulnerabilities in Classified US Government Systems
A U.S. official told the Associated Press that Anthropic’s Mythos AI model identified vulnerabilities in highly sensitive and secure U.S. government computer systems during a testing exercise. The official, speaking anonymously, stated that Anthropic collaborated with U.S. intelligence agencies to conduct the tests. Mythos identified certain vulnerabilities within hours, though the official noted that this does not mean the model could exploit them in that same timeframe. The revelation underscores the growing role of advanced AI in national security and cyber defense.
Anthropic’s Mythos model found vulnerabilities in classified US government systems, official says →