Open-weight models now hack nearly as well as restricted ones, and criminals are already running the agents.
Vals AI's network-intrusion test put a freely downloadable DeepSeek model within eight points of OpenAI's gated cyber model, at about a seventh of the cost per run, and Microsoft says a Cambodia-based group is pointing the ARTEX agent framework at banks. On the defensive side, the same week's tests show agents still finish only about a quarter of real coding requests in full. It is a lighter Sunday: 14 items, because weekend publishing is thin.
TLDR
- Vals AI's CyberOffense test puts open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ DeepSeek V4.1 Flash within eight points of OpenAI's restricted cyber model, at about $3 a run versus $22 (Nov 7). Security Open Source
- Microsoft says a Cambodia-based group, Storm-3358, is using the ARTEX agent framework against financial firms (Nov 6). Security Agents
- Co-Design Bench: the best of ten coding agents fully completes only 26.4% of real change requests (Nov 8). Research Agents
- OpenAI will remove gpt-5.5-cyber-preview from its API on February 1, 2027 (Nov 5). Pricing
- Meta's strategy chief Dave Wehner is leaving to become CFO of inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ platform Fireworks, which is reportedly in early talks at a $30 billion to $40 billion valuation (Nov 6). Business
Upcoming Deadlines
| Date | What | Who it hits | Do this | Source |
|---|---|---|---|---|
| Mon Nov 9 | SEC roundtable on machine-understandable data and "Super Intelligence" | Public-company and data teams | Watch for stream | SEC |
| Mon Nov 9 | GitHub Actions pull_request_target brownout, 16:00-18:00 UTC: blocked runs fail in public repos | Public repos with pull_request_target workflows | Check Insights for runs the default rule would block; allow the trigger in an Actions event policy or fix the workflow | GitHub |
| Tue Nov 10 | CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ federal patch deadline for Brotli CVE-2026-111497 (Node2Shell) | Anyone running Node.js, .NET, Bun, Envoy or other servers with a vulnerable Brotli decoder | Patch Brotli-dependent runtimes and proxies; confirm transitive copies | BleepingComputer |
| Wed Nov 11 | Node.js security releases for 26.x, 24.x and 22.x (highest severity HIGH) | Anyone running Node.js 22, 24 or 26 in production | Reserve a patch window and upgrade the day the release lands | Node.js |
| Thu Nov 12 | Rust 1.100 stabilizes Cargo minimum publish age (dependency cooldown, off by default; Socket, Oct 20) | Rust teams | Set a cooldown in Cargo config | Socket |
| ~Thu Nov 12 | OpenAI's proposed end of model access in Cursor (official date not confirmed; Cursor may end it sooner) | Cursor users on OpenAI models | Check Cursor terms | OpenAI |
| Mon Nov 16 | GitHub Copilot retires Claude Opus 4.8, Claude Opus 5, GPT-5.6 Terra (replacements Opus 5.5, GPT-6.1 Sol / GPT-6 Luna) | Copilot users with pinned models | Repoint model pins | GitHub |
| Mon Nov 16 | GitHub Actions pull_request_target brownout, 16:00-18:00 UTC: blocked runs fail in public repos | Public repos with pull_request_target workflows | Check Insights for runs the default rule would block; allow the trigger in an Actions event policy or fix the workflow | GitHub |
| Tue Nov 17 | Gemini API shutdown (earliest possible): gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-2.5-flash/pro TTS previews | TTS and Live API callers | Move to gemini-3.8-live or gemini-3.8-flash-tts | Gemini API |
| Tue Nov 17 | Google Gemini Gems become Skills | Gems users | Review Gems | |
| ~Thu Nov 19 | PostgreSQL 19 expected release (delayed, 8-10 features moved to 20; date not confirmed) | Postgres upgrade planners | Re-plan upgrade window | The Register |
| Fri Nov 20 | Meta updated commercial terms take effect (moved from Oct 30 after publisher objections; crawling clause now limited to advertising and business products) | Anyone with a Meta business/ads account tied to their domains | Re-read live terms; check robots.txt and staging logins | TechCrunch · Meta |
| Fri Nov 20 | Conductor Cloud shuts down after Meta acquisition (desktop app stops receiving new models in early 2027) | Teams running parallel coding agents on Conductor | Export work and move to another agent orchestrator | The Decoder |
| Fri Nov 20 | CISA KEV federal patch deadline for SonicWall SonicOS SSL-VPN SAML bypass CVE-2026-19868 (exploited) | Anyone running Gen7 SonicWall firewalls (TZ, NSa, NSsp, NSv) with SAML SSL-VPN login | Upgrade SonicOS to the fixed release; review SSL-VPN logins for the SAML path | BleepingComputer |
Action Items
- Patch every internet-facing web application with anonymous file upload or server-side request forgeryServer-side request forgeryA bug where an attacker makes a server send requests for them, reaching internal services it shouldn't expose.OWASP ↗ (SSRF) exposure this week; Microsoft says ARTEX operators scan for exactly those two flaw classes.
- Search your web servers for unexpected web shells and for netcat or curl downloads of remote-access tools, and load Microsoft's published ARTEX indicators into your detection tooling.
- Find every call to
gpt-5.5-cyber-previewin your code and repoint it to a current cyber model before February 1, 2027. - Budget a human cleanup pass on agent-written changes: do not count a coding-agent task as done until someone has checked each stated requirement, since the best model fully met only 26.4% of tasks in Co-Design Bench.
- Set your AI-auditor review threshold to a false-alarm rate your reviewers can absorb, and use auditors to estimate misbehavior rates rather than to gate releases.
Models & Tools
- GitHub Copilot code review now has REST and GraphQL settings APIs (Nov 6) Agents Enterprise administrators can enable or disable code review across organizations, and down to single repositories. Review comments now carry attribution fields, the effort level used, and the skills, MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ servers and custom instructions Copilot reported as sources. The review overview comment is more compact. Available on Pro, Pro+, Max, Business and Enterprise plans. (GitHub)
- AWS adds Kiro bots: named, long-running agents with their own memory (Nov 5) Agents Each bot starts with only the tools it was given, keeps its own memory, and launches and monitors its own parallel worker sessions, such as reading a ticket queue. AWS frames it as delegating processes instead of tasks. It builds on Kiro Crew, which already carried memory across sessions. (The New Stack)
- Meta adds Spaces to its Muse agents (Nov 6) Customer AI Spaces are topic-specific conversations that live in the Muse sidebar. Describe a tracker, planner or project board and Muse builds it, then keeps it updated through connected third-party apps in the background. The examples are consumer ones: trip plans, gift trackers, wedding planning. (TechCrunch)
- A homelab runs the 2.8-trillion-parameter Kimi K3 on used Optane memory and no GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ (Nov 8) Open Source The build uses a 40-core Xeon, second-generation Optane PMem 200 modules and 1-bit quantization, and manages 0.85 tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ per second. The builder calls active DIMM cooling mandatory. The smallest community quantization needs at least 610GB of RAM, so this is a capacity trick, not a speed one. (Tom's Hardware)
APIs & Pricing
- OpenAI removes
gpt-5.5-cyber-previewfrom the API on February 1, 2027 (Nov 5) Pricing The deprecation notice names no specific successor, only "the most capable cyber model available to you." It sits beside the already scheduled November 30 shutdowns of Evals, Agent Builder andv1/prompts. Cyber access is gated by account, so check what your organization can call before the cutoff. (OpenAI)
Security
- ARTEX is now in a financial-crime playbook, Microsoft says (Nov 6) Security Agents High · CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗: not applicable (no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗; the attacks use common web flaws) · Exploited in the wild · Patch: none, harden upload and SSRF paths. Storm-3358, a Cambodia-based cluster, began using the agent framework in October against financial firms, mostly in Southeast Asia. ARTEX agents scan, exploit anonymous file upload and SSRF, drop open-source RATs (WISP, vshell, AdaptixC2), and reuse stolen cloud identities. Scanners will not flag this as a vulnerability. (Microsoft)
Industry & Business
- Meta's strategy chief Dave Wehner leaves to be CFO of Fireworks (Nov 6) Business Fireworks runs and customizes open models for developers and was valued at $17.5 billion in a July round. Business Insider reported early talks at $30 billion to $40 billion, which is unconfirmed. A 14-year Meta finance and strategy veteran moving to an open-model inference host is a signal about where the money expects to sit. (CNBC | The Information)
- Two Chinese companies approved to buy Nvidia H200 chips have empty offices (Nov 8) Policy Business The New York Times, as relayed by Tom's Hardware, found that addresses for New Matrix and Maginfra led to empty offices, and that their executives link to firms sanctioned for supplying China's military. A former White House China director questions whether Commerce is checking license terms. Nvidia has not been quoted in the excerpt. (Tom's Hardware)
- Dutch police arrest 150 to 200 protesters at a Microsoft data center site (Nov 8) Policy The "Stop Dirty Data Centers" coalition says this is the first of actions planned across Europe, including Dublin and Warsaw. Police say some entered the Amsterdam construction site with violence. The campaign targets fossil-fuel power for data centers, so expect siting and power-source scrutiny to widen beyond the US. (Tom's Hardware)
AI in Production
- Aviator reviews 4x more pull requests by checking five things instead of reading every line (Nov 7) Production Agents The checks are intent, acceptance criteria, invariants, decisions and evidence. With nine engineers, weekly merged PRs rose from about 100 to over 400, median time to merge fell from 2.1 days to 3.4 hours, and reverts fell 40% (company-reported). A human still signs off. Worth stealing: let machines enforce repeated rules and keep people for decisions. (The New Stack)
Benchmarks & Evals
- Vals CyberOffense: open-weight DeepSeek V4.1 Flash 1017 reaches 72.8% of objectives against 80.8% for GPT-6 Sol Daybreak (Nov 7) Security Open Source Each agent starts with only a public web address. DeepSeek costs about $3 a run versus roughly $22, and made most progress only after 100 million tokens, so short tests understate open models. On the harder Lighthouse test, Daybreak completed 27 of 31 objectives and DeepSeek 21. (The Decoder)
- Co-Design Bench: GPT-6 Astra fully completes 26.4% of real change requests (Nov 8) Research Agents Alibaba-led researchers built tasks from actual coding sessions, favoring requests users had to rephrase or roll back. Across ten models, the best met about 70% of requirements on average, but usually left at least one detail unfinished. Older benchmarks top 99%, so they no longer separate models. (The Decoder)
- AI auditors catch most agent misbehavior but flood reviewers with false alarms (Nov 7) Research Agents Andy Hall, writing personally, replicated a METR study at a fixed 5% false-alarm rate. GPT-5 caught nearly 98% of reward-hacking cases, while Gemini 3.8 Flash found barely half. His conclusion: auditors suit estimating how often things go wrong, not gating on every incident. (The Decoder)
Paper Spotlight
- Richard Sutton says compute alone will not deliver intelligence (Nov 8) Research The author of "The Bitter Lesson" argues in a new post that AI science is "far from finished" and needs new methods. He repeats earlier criticism of synthetic data and of scaling language models alone, and argues for agents with world models that learn continuously from experience. Opinion, not a result, but it drew wide discussion. (The Decoder)