KD Agentic · AI Daily Digest · November 6, 2026
Seven stories from the past 24 hours: OpenAI turns its Preparedness Framework up to Critical with a dedicated security model, Meta buys an agent orchestration startup, Anthropic ships enterprise-grade security controls for Claude, and a new benchmark says AI can spot the variables but cannot write the law.
1. OpenAI Launches GPT-6 Cyber, Its First Critical-Rated Security Model
OpenAI has officially launched GPT-6 Cyber, a model purpose-built for security research, and rated it Critical under its own Preparedness Framework, one notch above the High rating held by GPT-5.6 Cyber. On ExploitGym the new model reaches a 48.5 percent success rate against 32.1 percent for its predecessor, a relative improvement of about 51 percent, and it completes 95 percent of tasks on RemoteExploitBench. In web penetration testing its vulnerability recall hits 40.3 percent, ahead of even the unrestricted GPT-6 Astra at 32.0 percent.
The headline results come from real codebases rather than benchmarks. In Firefox, GPT-6 Cyber found 81 previously unknown vulnerabilities that Mozilla classifies as high severity, plus evidence that 30 known issues were more serious than previously assessed. Mozilla's distinguished engineer Brian Grinstead noted that Firefox entered 2026 as one of the most heavily reviewed codebases in the industry, and the model still found issues that prior frontier models missed. Findings, reproductions, and candidate patches went to Mozilla, and some fixes have already shipped. In a separate November 4 research writeup titled "A thousand fewer paths to root", OpenAI ran an automated pipeline across the Linux kernel: 2,308 candidate issues, 1,724 dynamically reproduced, and 948 working local privilege escalation exploits, compared with 166 for GPT-5.6 Sol under the same framework. The whole run cost about $1.21 million in API credits, roughly $1,279 per working exploit, and 304 critical findings went to maintainers first.
The alignment work may matter more than the capability jump. In the ExploitGym Honeypot evaluation, where tempting out-of-scope targets are deliberately placed around the assigned task, GPT-6 Cyber stayed inside its authorization boundary 99.93 percent of the time without deployment safety prompts, up from 94.49 percent for GPT-5.6 Cyber, and 100 percent with prompts. Access is deliberately narrow: the model reaches vetted Daybreak Red customers first through API and Codex, with IP allowlisting on by default, while Daybreak Blue serves verified defensive researchers. The pattern is clear: OpenAI is not just shipping a stronger model, it is deciding who can touch which capabilities. — OpenAI · Sina Finance
🔗 OpenAI · Sina Finance
2. Meta Buys Agent Coding Startup Conductor
Meta has acquired Conductor, the startup behind a tool that lets engineers run multiple AI agents in parallel workspaces to build software projects. Meta confirmed the deal to TechCrunch but declined to share terms. Founder and CEO Charlie Holtz announced it himself on Wednesday: "Conductor is joining Meta", adding that the team has "big and crazy plans for the next interface" and promising something new for customers early next year.
Conductor was founded in 2024 and shipped its core product last summer. Its pitch is orchestration: agents deployed across cloud, local, or even phone-driven workspaces, each handling relatively independent modules of a project while staying coordinated. A detail worth noting is that the agents on the platform are not homegrown; they come from Anthropic, OpenAI, and other labs, which makes Conductor a bet on coordination layers rather than any single model. Holtz said he looks forward to working with Meta's Chief AI Officer Alexandr Wang and MSL product lead Nat Friedman.
There is a hard deadline buried in the announcement: Conductor's cloud service shuts down on November 20, after which the company begins deleting user cloud data, so existing customers need to migrate now. Core features keep running the old way until early next year. For Meta, the acquisition adds an orchestration team to Superintelligence Labs at the moment every lab is trying to turn chatbots into fleets of working agents. — Meta (confirmed via TechCrunch) · Founder announcement
🔗 TechCrunch · Conductor
3. Anthropic Ships Enterprise Security Controls for Claude
Anthropic published a security-focused update for Claude Enterprise on November 5, aimed at letting security teams see and govern what Claude does without rebuilding their tooling. The core piece is generally available now: prompt hooks and pre-tool-call hooks that send prompts and proposed tool calls to supported security vendors, so companies can apply policies they already wrote, including data loss prevention rules, at the moment a prompt is submitted or before a tool runs.
The rest lands in beta. A new security center compares chat, Cowork, and Claude Code settings against Anthropic's own recommendations, refreshes automatically, and can export auditor-ready reports; Anthropic is careful to say it offers guidance, not compliance certification. A parent console lets one organization manage SSO, directory sync, domain verification, and groups across multiple Claude Enterprise and Claude Platform organizations, with the caveat that setup requires approval from every involved admin and cannot be undone. A new Admin API endpoint can sign a member out everywhere in one call, ending sessions, trusted devices, and OAuth tokens without removing them from the organization.
The most infrastructural piece is detection-only: a supported vendor's sensor can run inside the isolated virtual machine that Cowork and unified Claude use for local code execution, reporting to the company's existing security console on macOS and Windows via device management. Each VM start verifies the sensor package's vendor signature. Anthropic is explicit that coverage stops at the local VM and does not extend to cloud Cowork sessions. Together the package reads as a response to a simple question security teams kept asking: if an agent acts on company devices, who watches it? — Anthropic
4. OpenAI's October Developer Update: 8x Codex, a Decisions API, and a New Terra Model
OpenAI published its October developer update roundup, and the through-line is speed and decision latency. GPT-6.1 Sol Ultrafast generates tokens up to 8 times faster than standard mode inside Codex, reaching about 300 tokens per second, and up to 6 times faster through the API. For coding agents that spend most of their wall-clock time emitting edits and tool calls, that is a direct cut in task latency rather than a benchmark curiosity.
The second piece is the Decisions API, built on GPT-6 Luna. Developers hand it text, images, or structured data along with predefined questions and options, and it returns fast judgments for content classification, request routing, task prioritization, or agent action selection. OpenAI says it runs about 10 times faster than the traditional Responses API for these tasks. It opened on October 22 and has since gained prompt caching, JSON structured input, and a jump from 200 to 400 questions per request. Alongside it, the Responses API gained a 12 hour prompt cache option, and API usage tiers were reorganized into Build, Launch, and Grow.
On the model side, GPT-6.1 Terra targets visual reasoning, computer operation, and text processing, aimed at agents that navigate software interfaces at lower cost. Codex itself learned to predict your next instruction from context, accept voice-started CLI tasks, and run inside a new Windows sandbox that constrains network and filesystem access. Eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits. The stack reads as one bet: agents should decide faster, act cheaper, and stay contained. — OpenAI (October developer update) · Ifeng Tech
🔗 OpenAI · Ifeng Tech
5. SambaNova's Inference Chips Ride Into Supermicro Racks
Supermicro will build and sell AI inference systems that pair SambaNova's SN50 chips with NVIDIA GPUs, the companies announced on Tuesday, with an initial plan of 10,000 SN50 chips. The division of labor is unusually explicit: NVIDIA GPUs handle the reading of the prompt, and SN50 chips handle the writing of the answer, token by token. The racks join Supermicro's Data Center Building Block Solutions, with Supermicro handling assembly, testing, delivery, and on-site service while SambaNova engineers tune models on the hardware.
SN50 is SambaNova's fifth-generation Reconfigurable Dataflow Unit, and its pitch is fit rather than raw density. A SambaRack SN50 draws an average of 20 kW, which means operators can add inference capacity inside data centers they already run instead of waiting for new builds constrained by power and cooling, and keep existing GPUs in service longer. First targets are neoclouds and sovereign clouds, followed by enterprises running AI in their own facilities. "Agentic AI is the next wave of inference, and it has to run everywhere," said SambaNova CEO Rodrigo Liang.
The deal fits a busy quarter for the San Jose company: Oracle said on October 27 it will deploy SN50 chips in its cloud, services firm Emergys signed on October 29, and SambaNova raised at an $11 billion valuation in July with Foxconn and Intel joining a rackscale effort in June. Inference is becoming a layered market, and the decode stage just got its own distribution channel. — SambaNova · Supermicro
🔗 SambaNova · Supermicro
6. Unsatz Benchmark: AI Agents Find the Variables but Not the Law
A new arXiv paper introduces Unsatz, a benchmark that tests whether LLM agents can do the hardest part of scientific discovery: inventing the mathematical structure of a hypothesis, the ansatz, before any coefficient is fitted. Earlier equation-discovery benchmarks hand the model a variable list and candidate terms on a plate. Unsatz does not. Across 80 simulated systems spanning 20 scientific domains, from delayed differential equations to causal models and probabilistic processes, agents must buy experiments within a fixed budget, write candidate programs in a miniature modeling language, and are scored purely on predicting experiments they have never seen, including interventions and conditions outside the explored range.
The results are sobering. Nine language models solved between 21 and 60 percent of the systems. Strikingly, classical equation-discovery methods failed completely on 44 of the 80 systems even when the true equation form was provided with only the coefficients unknown, confirming that finding the structure, not estimating the numbers, is the real bottleneck. Behavioral analysis adds a sharper edge: stronger models make fewer unscientific errors, but models broadly succeed at identifying which variables exist while failing to grasp how those variables physically interact, and they prefer revisions that lower training error over revisions that improve prediction on unseen data.
For anyone running AI-for-science programs, the practical lesson is blunt. Feeding an agent a program discovered for a similar system only helped when the underlying mechanism was shared and the first investigation succeeded. Until models learn to resist the lure of number-matching, human researchers should keep ownership of the structural hypothesis and let agents explore within it. — arXiv
🔗 arXiv
7. Anthropic Extends Text Watermarks to Older Claude Models
Anthropic is expanding its invisible text watermark to older Claude models, per a second-round notice to API customers reported on November 5. The first wave went live on November 3: text from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 now carries the mark, according to Anthropic's official release notes. From November 16, four more models join: Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5, with Amazon Bedrock rollouts a few days behind. Opus 4.6 was originally scheduled for November 3 and slipped without explanation.
The catch sits in the sampling parameters. For all four models in the second wave, the watermark applies only when temperature, top_p, and top_k stay at their defaults; change any one and the output may come out unmarked. Anthropic warns customers who rely on the watermark for compliance to keep these settings at defaults. In the first wave, only Opus 4.6 and Sonnet 4.6 carried that restriction.
The why is regulatory rather than technical. Anthropic signed the EU AI Act's Article 50(2) code of practice on AI-generated content transparency in July alongside roughly 190 other signatories, and systems placed on the market before August 2, 2026 fall under a transition period that closes around December 2. The watermark itself is a variant of Google DeepMind's SynthID-Text method, embedded at the model level across AWS, Google Cloud, and Microsoft Foundry, invisible to readers, and not disableable by customers. OpenAI took a narrower path with its textGrain mark in October, default-on only in the EU. How regulators treat conditional marks, where one sampling flag removes the signal, may become the compliance baseline question of December. — Anthropic release notes · The Decoder
🔗 Anthropic · The Decoder
KD Agentic · AI Daily Digest, November 6, 2026
