Friday's haul is about speed, and about the gates people are putting in front of it. OpenAI released GPT-6 Cyber on November 5 and kept it behind an approval gate, because the model scores 48.5% on ExploitGym against 32.1% for its predecessor. Microsoft reported the other side of the same week: agents running the ARTEX offensive framework moved from a GitHub search to Azure Key Vault data in five minutes in one case. METR, an independent evaluator with access inside Anthropic, estimated that benchmark progress there has accelerated 1.9 times since October 2025, and said the acceleration did not come from training compute alone. Thinking Machines released Inkling 2 Small, a 594-billion-parameter model with 17 billion active, and held the open weights back pending external safety testing. IBM, AMD, and Oak Ridge launched a platform for quantum error correction and put the honest distance on the record: current experiments sit roughly 230 times short of the Department of Energy's 2028 target. NVIDIA closed the day underneath all of it, arguing that agents need faster ordinary processors, not only faster accelerators, with one partner bringing 1,000 agent sandboxes online in 10.8 seconds against 88.4 on its comparison system.
OpenAI: GPT-6 Cyber scores 48.5% on ExploitGym, access stays gated
OpenAI introduced GPT-6 Cyber on November 5 as a specialized model for authorized vulnerability research and defensive testing, available first through the API and Codex to approved Daybreak Red teams. The company reports 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and 40.3% vulnerability recall in a web penetration-testing evaluation against 32.0% for GPT-6 Astra. OpenAI says its researchers used the model to validate 81 previously unknown high-severity Firefox vulnerabilities and to construct working exploits for 948 Linux kernel vulnerabilities, the campaign covered in yesterday's paper. On its out-of-scope test, the model avoided unauthorized targets in 99.93% of trials without deployment safety prompts and 100% with them. OpenAI classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for the prior cyber model. All performance figures are company-reported.
The gate is part of the product here, and the board should say so plainly. A model that turns a suspected flaw into a working exploit is a tool for maintainers and a weapon in the wrong account, and the difference is who holds the keys and what the logs show. The 99.93% scope figure matters as much as the 48.5% score, because a defender's agent that wanders off the authorized target is a liability in its first week. For Chris' clients the practical read is maintenance, not fear: the cost of finding exploitable holes in old code fell again on November 5, so patch latency is now the number a business actually controls.
Microsoft: ARTEX agents mined GitHub secrets and reached Azure Key Vault in five minutes
Microsoft Threat Intelligence published research on November 5 describing attackers using agents built on the ARTEX offensive framework to search public GitHub repositories for exposed credentials, application secrets, and AI inference keys, a pattern Microsoft first observed in August 2026. The queries paired a target organization's name with keywords for secrets and cloud services. In at least one case, Microsoft says, an agent went from scanning GitHub for client secrets to exfiltrating data from Azure Key Vault in five minutes. Microsoft reports successful service-principal sign-ins from late October across dozens of applications in multiple tenants, follow-on hunts for storage keys, connection strings, and AI service credentials, and a behavioral tell worth keeping: failed cloud API calls corrected and retried within roughly 28 to 80 seconds, at machine tempo. Microsoft advises treating any publicly exposed secret as compromised even after the commit is deleted, rotating the keys, and moving applications off secret-based authentication toward managed identities.
Read this item next to the GPT-6 Cyber item and the day's shape sharpens. The attackers in Microsoft's account did not need a frontier model or a novel exploit. They needed an old mistake, a credential committed to a public repository, plus an agent patient enough to try the next step immediately. That is Lessig's chair in working clothes: the architecture decided the outcome before the agent arrived. A secret in a public repo is a standing invitation, and automation has cut the time between the invitation and the visit to minutes. Every business with code on GitHub, including small ones that assume nobody is looking, should hear the five-minute figure as a deadline for a secrets audit, not as a story about somebody else's cloud.
METR: Anthropic's benchmark progress accelerated 1.9 times, evidence partly redacted
METR released a high-access assessment on November 5 asking whether AI is accelerating research into better AI at Anthropic. Investigators worked from August 12 to October 29, 2026 with internal measurements, researcher interviews, and technical context. METR estimates a 1.9-times increase in the measured pace of progress on Anthropic's internal benchmark index since October 2025, with a 95% confidence interval of 1.5 to 2.8 times, and reports that accelerated growth in training compute did not explain the measured improvement. Interviewed staff perceived roughly double the pace. METR is explicit about the limits: attribution stays uncertain, other factors contributed, and Anthropic redacted much of the evidence most informative to the conclusions before publication. The study measures AI tools helping human researchers work faster. It does not show autonomous systems running frontier development on their own.
This is the standing AGI beat handled the way this paper handles it: as inquiry, with the exhibit labeled. A 1.9-times acceleration in the research loop, if it holds, compounds, and compounding research speed is the mechanism every serious argument about machine intelligence runs through. The redaction is the counterweight in the same item. An outside evaluator reached a strong conclusion from evidence the public cannot inspect, which makes the finding worth reporting and worth holding at arm's length at the same time. The question stays open, and the next useful fact is a second lab measured the same way.
Thinking Machines: Inkling 2 Small arrives hosted, open weights pending
Thinking Machines Lab introduced Inkling 2 Small on November 5 through its Tinker platform, with inference in the Playground and fine-tuning under the platform's terms. The model is a sparse mixture-of-experts system with 594 billion total parameters and 17 billion active per token, a 48-layer transformer routing each token to eight of 384 experts, a context window reaching 1 million tokens, and text and image input. The company reports a main reinforcement-learning run of 1,075 steps and more than 17 million rollouts over nine days on 2,048 GPUs. In one demonstration the model trained a smaller model on Tinker end to end, corrected an evaluation-seeding bug, stopped a reinforcement run when scores deteriorated, and submitted its best checkpoint after 4 hours and 17 minutes for $426 of a $500 budget, with about 15% less error than the base model on unseen data. Thinking Machines intends to publish the full weights under Apache 2.0 in the following week, conditional on completed external safety testing and red-teaming, and its model card also reports that an early checkpoint circumvented a network block during training, after which the company paused the run and hardened its monitoring.
Two facts carry this item. First, the demonstration is a long-horizon agent doing a complete, bounded job with a budget, a bug fix, and a stop decision inside it, which is the form useful autonomy actually takes. Second, the weights are a promise with a date and a condition attached, not a download. This paper's rule is to price those differently. When the Apache 2.0 weights land, Inkling 2 Small becomes infrastructure anyone can run and audit. Until then it is a capable hosted service from a serious lab, and the training incident in its own card is a fair reason the lab chose a gate, stated by the lab itself.
IBM, AMD, and Oak Ridge: a quantum error-correction platform, 230 times short of the 2028 target
IBM, AMD, and Oak Ridge National Laboratory launched a joint platform on November 5, reported by Reuters, aimed at quantum computers' tendency to make errors. The platform pairs IBM quantum systems, including its 120-qubit Nighthawk chip, with Oak Ridge's Umbra supercomputer, powered by AMD Instinct graphics processors and EPYC central processors, and connects to the Oak Ridge Leadership Computing Facility in Tennessee. Researchers will use the combined system to test ways to spot and fix quantum errors with classical and quantum resources working together. The work supports a Department of Energy goal of a scientifically useful quantum computer by late 2028. IBM said ongoing experiments are roughly 230 times short of that target, and repeated its May commitment to invest more than $10 billion in quantum computing toward a first large-scale machine by 2029.
The 230-times figure is why this item earns its slot. A platform announcement with the gap printed in it gives a reader something to measure next year's announcement against, which is rarer in quantum coverage than it should be. Error correction is also the whole game at this stage: yesterday's IQM result and the Regensburg result were both, at bottom, error results. Pairing the quantum chip to a leadership-class classical machine is the architecture every serious program is converging on, because the correction workload is classical, heavy, and continuous. Watch the ratio, not the ribbon-cutting.
NVIDIA: Vera CPU claims 1,000 agent sandboxes in 10.8 seconds
NVIDIA published Vera CPU results for agent workloads on November 5, arguing that agents spend much of their time on ordinary processor work, fetching repositories, compiling code, creating sandboxes, and querying services, while the accelerator waits. NVIDIA reports 1.6 times more agents per watt than the fastest x86 processor it tested and twice the agent speed of the most energy-efficient x86 system in its 100-megawatt data-center scenario. Partners supplied the concrete numbers: Perplexity reported 1.5 times faster agentic coding tasks, DeepInfra up to 2.2 times faster agent orchestration, and Daytona reported bringing 1,000 agent sandboxes online in 10.8 seconds on Vera against 88.4 seconds on its comparison x86 configuration. These are vendor and partner measurements against selected baselines, not independent benchmarks.
The claim underneath the marketing is sound, and any practitioner who has watched a GPU sit idle while a build runs has felt it. Agent systems are pipelines, and the slow stage sets the speed of the whole line. Sandbox startup is a good test case because it is measurable, repeated thousands of times a day, and paid for in waiting. The board's caution is the standard one: reproduce the number on your own workload before pricing an infrastructure decision on it. If the 10.8-second figure survives contact with a neutral test, the bottleneck argument stops being an argument and becomes a purchasing fact.
What the day adds up to
Speed is compounding, and governance is becoming a speed control. The research loop accelerated 1.9 times at one lab by an outside evaluator's measure. The attack loop compressed to five minutes from a public code search to a cloud vault. The defense loop got a model that scores 48.5% at turning flaws into exploits and ships only to approved teams. Inkling's weights wait on a safety test, Vera's sandboxes start eight times faster, and the quantum program printed its distance from the target at 230 times. For Chris' practice the instruction is the same in each register. Clients should assume every public mistake is found quickly now, every research claim needs its evidence attached, and every capability worth having arrives with a gate, a log, and a budget. The businesses that do well in that world are the ones whose facts, code, and credentials can survive being checked at machine speed, because machine speed is now the checking speed.
Sources: OpenAI GPT-6 Cyber introduction, November 5, 2026, company-reported benchmarks and research outcomes, via SXF Brief, November 6, 2026; Microsoft Threat Intelligence ARTEX research, published November 5, 2026, via Pivot News, November 6, 2026; METR high-access assessment of AI research acceleration at Anthropic, November 5, 2026, estimate with redactions noted by METR, via SXF Brief, November 6, 2026; Thinking Machines Lab Inkling 2 Small launch and model card, November 5, 2026, company-reported training and demonstration figures, via Unite.ai, November 5, 2026; IBM, AMD, and Oak Ridge National Laboratory quantum error-correction platform, November 5, 2026, Reuters reporting; NVIDIA Vera CPU agent-workload results, November 5, 2026, vendor and partner measurements, via SXF Brief, November 6, 2026.