Tap to play. It never plays by itself, and the captions are in the picture.
Microsoft says your AI agents need watching like insiders. Microsoft launched Guardian Agents to watch a company's AI agents. It says Defender now spots up to 60 agent types. Attackers hit OpenCode Console, where some users kept API keys. TechCrunch says 1.4 percent of users were affected. Nvidia released Nemotron 3.5 Super VL, a 120 billion parameter open model for agents. In Anthropic's survey, staff expect AI to do 75 percent of their work by 2030. US workers say 28. Inception says Mercury 3 runs at 1,068 tokens per second. Scope agent access, and keep a person on risky actions.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
TechCrunch reports Microsoft launched Guardian Agents to monitor a company's AI agents, and Defender can now find up to 60 types of agents employees bring in, including Claude Code and Codex.
Why it matters: Microsoft's security chief says frontier models cannot be trusted. Its new tools limit agent access, adjust permissions by task and can require human approval for risky moves.
TechCrunch says attackers reached an older OpenCode Console system where some users stored API keys, hitting 1.4% of users. Anthropic and Google blocked exposed keys; OpenAI refunded credits used.
Why it matters: Co-founder Dax Raad says AI models stopped helping analyze the attack when malicious code tripped safety filters. Keys kept in hosted consoles are a target.
TechCrunch reports Super VL is a 120-billion-parameter open-weight model for enterprise agents, plus Bolt models for local PCs. Super VL and Bolt 74B-A9B are available now; Bolt E4B later this month.
Why it matters: Nvidia names Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce as Super VL users. Teams can post-train it on their own data and run it on their own hardware.
Anthropic surveyed US adults and its own staff in August 2026. 91% of staff vs 39% of US adults expect AI to make a Nobel-level discovery by 2030. Both groups tie own-job exposure to higher unemployment.
Why it matters: Workers who expect AI to do 80% or more of their work forecast 12.2% unemployment in 2030, Anthropic reports. The groups agree on today's AI but split on the pace.
Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks while running 4–8x faster, with a 1M-token context. Launch price is $0.10 per 1M input, $0.35 output.
Why it matters: It's on OpenRouter, Baseten and Vercel and is OpenAI API compatible: a fast, cheap option for latency-bound agents. The quality claims are Inception's own.
TechCrunch reports TorchNeuron lets PyTorch apps run on Amazon's Trainium chip by changing one line of configuration. Amazon says about 85% of AI apps use PyTorch and it is upstreaming changes.
Why it matters: It's AWS's most direct push against Nvidia's software lock-in. AWS says Deepgram processes 70% more audio per dollar after moving to Trainium.
AWS says Bedrock now enforces JSON schemas during generation for Claude on both the bedrock-runtime and bedrock-mantle endpoints, and strict tool use makes tool inputs match their schema.
Why it matters: Fewer agent failures from malformed tool input. AWS notes refusals and token limits can still break the schema, and number and string length limits are not enforced.
Researchers report HexaShell, a thin harness where Codex with GPT-6-Astra writes and fixes its own tools after failures, reached 39.7% success across all 42 RoboDojo tasks, a new state of the art.
Why it matters: The limit was contact, not reasoning, the authors say: pushing hit 92.5% but tight insertion only 8.3%. The agent usually saw its failures but couldn't make the motion work.
Researchers report Self-HPO alternates harness changes the model proposes from its own failures with RL on the model. Across three coding benchmarks it raised Qwen3.5-35B-A3B's average resolve rate.
Why it matters: The authors say the joint gain beat the sum of harness-only and policy-only gains: a data point for teams deciding whether to tune the harness or the model.
On the Frontier watch →Databricks says HP uses Genie and agent-based tools on its Marketing Lakehouse. HP estimates up to 40% lower CDP cost of ownership and cut governed data-sharing from three months to 48 hours.
Why it matters: HP's lesson, per Databricks: agents work reliably only once data, business definitions and governance are unified. These are early results reported by the vendor.
MIT Technology Review says Meta's Muse was downloaded 5 million times in its first month, while sites like Yelp and Amazon block agents and Shopify gives them their own checkout lanes.
Why it matters: Whether sites block agents or redesign for them decides if consumer agents work at all, the piece argues, and a web built for agents could get worse for humans.
TechCrunch, citing Bloomberg, says the two are on the verge of another $45 billion compute deal, a little over a month after their first. Sources caution talks are ongoing and details could change.
Why it matters: Compute commitments keep growing as both firms pursue IPOs; Bloomberg reports investor pushback on Nscale's proposed $30 billion valuation.
CNBC says Nvidia named Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace as targets, without sizes, and signed a declaration of intent with Germany's government on industrial AI models.
Why it matters: Siemens says AI tools built with Nvidia could cut production engineering work by up to 85%. Merz asked smaller firms to open their data to AI.
TechCrunch reports Mamdani spoke at an AI Now Institute event, after ordering an AI ban in elementary and middle schools last month and proposing AI disclosure for property listing photos.
Why it matters: With little federal AI regulation, TechCrunch notes, cities are acting on their own; New York's moves touch edtech and real estate listings.
At TechCrunch Disrupt, Irene Solaiman said evaluations are an evolving science, reproducing them needs model access, and agentic evals are hard and expensive, with few evaluations for harnesses.
Why it matters: Labs are leaning on third-party evaluators instead of regulation. Solaiman says the qualifications for an effective evaluator need much more standardization.
The repo describes itself as ears for coding agents: CPU-only and open source under MIT. It gained 312 stars since our last check; it was created on 2026-10-15.
Why it matters: The fastest-growing repo we track this run, adding audio as a sense for coding agents.
The repo describes an agentic coding tool that understands your codebase, runs routine tasks and handles git workflows. It added 261 stars since our last check; no license is listed.
Why it matters: The second-largest star gain we saw this run, on a repo created on 2025-02-22.
Two Hacker News threads on a DHS order banning DeepSeek from federal civilian agencies drew 65 and 43 points. @AndrewCurran_ posted that it also covers third-party services built on DeepSeek.
Why it matters: If the reports hold, vendors selling to US agencies may need to check their stacks for DeepSeek models. @ZackKorman posted that the order is absurd.
Benjamin Congdon writes that his prototype agent plans and investigates change-point monitors, and that having the agent write specs run by a deterministic system gave cleaner signals with fewer tokens.
Why it matters: He estimates a moderately large company could monitor its fleet for about $100/day with small models: one practitioner's estimate, not a study.