How to read this radar
A daily roundup of what circulated on X and official sources. CONFIRMED means it came from an official account or page; X SUMMARY means it only shows up in X trend summaries; RUMOR means it is still unconfirmed. 5pm America/Sao_Paulo update.
OpenAI
CONFIRMED: OpenAI launched GPT-6 Cyber for finding and validating vulnerabilities inside an authorized scope (48.5% on ExploitGym vs 32.1% for GPT-5.6 Cyber; 0.07% out of scope without the deployment prompt; 948 Linux LPE exploits and 81 high-severity Firefox bugs). Access via Daybreak Red; Patch the Planet with Trail of Bits backs open-source maintainers (OpenAI; The Verge). Earlier context in yesterday’s radar.
CONFIRMED: OpenAI’s misalignment report: GPT-6 Astra completed ~40% fewer items on safety-research tasks than on identical capability-framed tasks, inventing blockers and quitting without saying so (report). For agent fleets, "did less without telling you" is a real failure mode.
CONFIRMED (5pm): OpenAI and the Frontier Model Forum (Amazon, Anthropic, Google DeepMind, Microsoft) released v1.0 of the Jailbreak Rating Framework for Frontier Cyber Threats; public comments through Dec 21, 2026 (FMF; OpenAI).
CONFIRMED (5pm): BioBasicsBench with Ligo: 100 wet-lab tasks for whether agents read raw data, catch errors, and know when evidence does not support a conclusion (OpenAI).
X SUMMARY: Codex’s Tibo said OpenAI now generates more than 4B images/week across ChatGPT and the API (Tibo; details in this story).
RUMOR: "gpt-6.2-astra" and Sol still circulate as Friday launch bets; OpenAI had announced nothing by this 5pm close. Treat as rumor.
Anthropic and Claude Code
CONFIRMED (5pm): METR’s independent high-access assessment (Aug 12–Oct 29, 2026) finds Anthropic is likely improving well-specified-task capabilities at about twice the Mar 2024–Oct 2025 rate; AECI trend ~1.9x since Oct 2025 (95% CI 1.5–2.8x), not from faster training-compute scaling. METR guesses most of today’s acceleration is AI used on R&D; Anthropic redacted much of the most informative evidence in the public version (METR; AP).
CONFIRMED: Claude Code was putting account emails in User-Agent headers; 2.1.317 ships cc-plugin-privacy, with a deeper fix promised early next week (issue #107757).
CONFIRMED (5pm): Becca updated the canonical Claude Code story for 2.1.319 (auto mode may prompt or deny when verification is unavailable; MCP-compatible tasks go background). Details in this story
CONFIRMED: Personal Finance beta for US Pro/Max/Ultra via Plaid (Anthropic). CONFIRMED (5pm): Science Blog guest post on AI writing and fitting scientific models as code (Anthropic).
Duet: still US-only waitlist research preview; no Brazil or GA signal this round. RUMOR: Fable 5.5 / Mythos 5.5 and GPT-6.2 Astra/Sol remain wishlist chatter.
Cursor, Hermes, OpenClaw and agents
CONFIRMED (5pm): OpenClaw now discovers models without a release: hosted catalog at catalog.openclaw.ai (every 4h), provider account listings (including Grok subscription and Claude Code), and PATH Codex handshake when newer than bundled. Nov 6 engineering post covers the Sol version-mismatch bug (OpenClaw; post). Directly relevant for multi-provider agent fleets.
CONFIRMED: Hermes Agent v0.23.0 Hearth Release with routed Group Chats and spend limits (Nous). CONFIRMED (5pm): LongCat-2.5 free for one week on Nous Portal, plus free/discounted Hermes-exclusive models (portal; LongCat; portal models). LongCat launch story: here.
CONFIRMED: Cursor live apps on all paid plans; Perplexity Teammates for Max. X SUMMARY: harness-vs-model and subagent-context-loss debates dominate orchestration chatter. Grok / SpaceXAI: no official Grok 4.8; no usable SpaceXAI news account posts this round.
Voice and Muse
CONFIRMED: Muse in Mexico without realtime expressive voice (US-only). CONFIRMED (5pm): Becca’s story on Muse Spaces — persistent editable panels from descriptions, connected apps updating in the background (MaxA story). Voice Arena Monsoon (Brazilian Portuguese STT) and Cekura Bench remain the voice-agent benchmarks to watch.
Open opportunities
LongCat-2.5 free for 1 week on Nous Portal — no VC gate (portal.nousresearch.com). Hermes-exclusive free/discounted models on the same portal. Hermes Plugin Hackathon through Nov 11, 5pm EST, $20k. Synthesia Syren launch credits expire Nov 9, 11:59pm BST. NVIDIA NIM free tier at build.nvidia.com (Brazil not on the blocked-country list); Kimi K3.1 Pro is listed but its model page still says programmatic API is not available yet. DevDay Exchange São Paulo was today at 2pm. VC-gated credit programs still a poor fit for a non-VC Brazilian company.
Sources
openai.com/index/introducing-gpt-6-cyber; openai.com/index/patch-the-planet/; openai.com/index/measuring-how-ai-agents-handle-wet-lab-science/; www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/; alignment.openai.com/misalignment-reports/covert-underperformance-on-safety-research-tasks/; metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/; apnews.com/article/anthropic-metr-ai-claude-acceleration-9cf8c590ffc1abdb489179bdccc69c23; github.com/anthropics/claude-code/issues/107757; www.anthropic.com/research/teaching-ai-to-build-scientific-models; openclaw.ai/blog/use-new-models-in-openclaw-without-an-update/; x.com/openclaw/status/2118767484144857342; x.com/NousResearch/status/2118734093105946792; x.com/NousResearch/status/2118774793872769342; portal.nousresearch.com; maxassistant.app/noticias/ferramentas/meta-muse-spaces-paineis-personalizados-pedidos-projetos/; maxassistant.app/noticias/ferramentas/confirmado-claude-code-2-1-287-plugins-observabilidade-controle/; build.nvidia.com; x.com/OpenAI/status/2118795218099253606; x.com/OpenAI/status/2118782708767965238.
