PJ AI Newsletter: Week of Nov 2-6, 2026

On Nov 5, OpenAI introduced GPT-6 Cyber, a cybersecurity model for authorized vulnerability research, available first to approved Daybreak Red customers in the API and Codex. OpenAI rates it Critical in cybersecurity under its Preparedness Framework (up from High for GPT-5.6 Cyber) and reports 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber. The day before, in "A thousand fewer paths to root," OpenAI described a mid-October pipeline run that dynamically reproduced 1,724 Linux kernel vulnerabilities and produced 948 working local privilege-escalation exploits at an API cost of $1,213,430 (about $1,279 per LPE). The same campaign produced evidence for 81 previously unknown high-severity Firefox bugs, and Mozilla has been landing fixes. Access is gated; mainline models still cover everyday code review and triage.

On Nov 2, Anthropic released FlashProve, an open-source zkSNARK research prototype that lets a provider prove a response came from a specific committed set of model weights without revealing those weights. On Llama 4 Maverick (400B MoE, 4-bit), Anthropic reports about 110 tokens/second proving on one H100, roughly 19 MB proofs per 1,024-token request, and verification in about 17 seconds on a 24-core CPU. Overhead versus ordinary inference is roughly 70 to 95 times GPU time. The release includes code, a technical report, Lean proofs, and Apache 2.0 licensing. Anthropic is clear this is intermediate research: 1,024-token windows, integer/greedy decoding, architecture not hidden, and no formal guarantee that the fast C++ verifier matches the Lean proofs.

On Nov 5, Cursor launched Build Live Apps: describe an internal tool, and a cloud agent can scaffold the app, stand up Postgres and auth behind Cursor SSO, run security checks, and publish a shareable link. Team admins can see apps and secrets, add review rules, and unpublish. It is available on paid plans with cloud agents enabled. On Nov 6, Cursor followed with realtime Voice Chat with Agents across the Agents Window, IDE, web, and Cursor for iOS, powered by Grok Voice (including with Privacy Mode / zero data retention). You can talk while a coding agent works and hear short progress notes instead of reading every reply.

On Nov 4, OpenAI announced it will acquire Pierre Computer Company, the team behind Code Storage and open-source Trees and Diffs libraries already used in the Codex app for browsing files and reviewing changes. OpenAI says Pierre's tools will help adapt Git workflows to how teams work with coding agents, and that Pierre's open-source projects will stay open source. Closing is subject to customary conditions; until then the companies remain separate.

On Nov 5, Anthropic announced new Claude Enterprise security and admin features: a security center (beta) that compares chat, Cowork, and Claude Code settings with Anthropic's recommended configuration; inference hooks now generally available that send prompts and tool calls to your own data-security tools for allow/deny; centralized management at admin.claude.com; Admin API sign-out of a member; self-serve IP allowlists; and controls to block personal Claude accounts in Claude Desktop on managed computers. An EDR sensor can also run inside the local Cowork/Claude VM (beta).

On Nov 4, OpenAI released Images 2.6 in ChatGPT, Codex, and the API. OpenAI says generation is about 50% faster, and GIFs can be created from a text prompt or a photo in one step. The API adds GPT-Image-2.6 Flare and Sunburst, plus image markup, SVG export, and a complexity parameter. Images 2.6 is available on all ChatGPT and Codex subscription tiers; Flare can generate in as little as six seconds at low quality, per OpenAI.

On Nov 5, Google launched Gemini Omni 1.2 Flash as generally available on the Gemini API and Google Cloud. Clips go up to 15 seconds (from 10), with multi-turn conversational editing, stronger reference adherence, and a 30% price cut to $0.07 per second for 720p. You can draft at lower resolution, edit in conversation, then upscale. Outputs carry SynthID watermarks. Google Vids can generate 15-second clips at no additional cost with a Google account or Workspace subscription.

On Nov 2, Anthropic put message threads into public beta on the Messages API (and on Claude Platform on AWS, Microsoft Foundry, and Anthropic-managed Claude on Google Cloud). Agents send the full conversation once, then only what is new on later turns, while Claude still reads the full history. For most Claude Code users it is on by default: Anthropic reports about 35% lower median time-to-first-token on requests with 500k to 1M tokens. Threads sit encrypted in memory, expire after 5 minutes idle, and token usage is unchanged. Best for long coding agents, computer-use loops, and image-heavy sessions.

On Nov 5 at KubeCon week, Google Cloud said it is donating Agent Substrate 1.0 and gVisor to the CNCF (Sandbox), alongside OpenRL. Agent Substrate treats agent sessions as Actors that hold memory and disk state, multiplexed onto warm Worker pools, with claimed sub-500ms restore. That targets the idle problem: agents wait most of the time on model calls or humans, but you cannot kill the Pod without losing Git state and tool sessions. gVisor isolates untrusted model-generated code from the host kernel. Docs and code are at ate.dev and github.com/agent-substrate/substrate.

On Nov 4, OpenAI made personal "dots" (persistent agents that keep working between conversations) generally available to eligible Enterprise, Edu, and Healthcare workspaces. Admins get a preview window through Nov 18, 2026, then dots can turn on by default unless they opt out. GPT-6.1 Terra can power dots. The same week brought iOS notification replies, desktop browser/file annotations, and stronger Windows local-thread behavior. Residency and HIPAA eligibility apply where the workspace already supports them.