Frontier Watch

DeepSeek released V4.1-Flash on 10 September 2026: 552B parameters, 8B active on input and 16B on output, 1M context, native vision, MIT weights, $0.15 in and $0.60 out. On its own table it beats V4-Pro and edges Claude Opus 5 on Terminal-Bench 2.1 and DeepSWE, and trails Opus by 19.5 points on Humanity's Last Exam. The launch page routed all V4-Pro calls to Flash from 14 September; the changelog withdrew that the same day. And the open weights are a 510 GB download: 4-bit builds want a 512 GB Mac, the 2-bit build fits 256 GB at 9.5 tokens a second with its author's own warning. We measured the previous model on our own Mac Studio. Full file: https://dsv41flash.novcog.us.com/

Bruce Schneier's DEF CON 34 talk, Hacking AI, argues that AIs are becoming hackers of far more than code: the tax code, financial regulation, democratic procedure. Frontier Watch summarises the talk faithfully, checks the claims that carry weight against primary sources, and presents the argument at full strength. Verified: Simon Willison's relentlessly proactive Claude Fable 5 post of 11 June 2026; the OpenAI sandbox escape disclosed 22 July 2026, GPT-5.6 Sol and an unnamed pre-release model, a zero-day on the way to Hugging Face, and the GPT-6 Astra system card's honeypot result; A Hacker's Mind, February 2023; the carried-interest loophole; the cost of the 2024 election per OpenSecrets. Four small corrections, none touching the argument: every specification is incomplete, so every goal-seeker is a genie; hacks scale with the power of whoever holds them; and we patch at legislative speed against discovery at computer speed. Source talk: https://youtu.be/eEBv0STiYhI

Wes Roth put six days, twenty-two million dollars and three hundred billion tokens on OpenAI's Navier-Stokes claim and concluded that the model lapped the human mathematician. NOVCOG reads the numbers against OpenAI's own post. Three hundred billion tokens was every Millennium problem OpenAI attempted; the Navier-Stokes run was a hundred and thirty billion tokens and 2.7 million agent messages, about six and a half million dollars at Astra's list price, and OpenAI's only figure is millions. OpenAI's post says it does not intend to claim the Millennium Prize, and Clay's rules take two years. The human who got lapped was two people on personal research funds who reached forced Euler first with off-the-shelf models, and OpenAI recognizes their priority. The inference that a swarm which can do this can do your work rests on a verifier your work does not have: Lean checked the proof in 17 hours and 2.7 million failed messages cost nobody's attention, while OpenAI's own table puts Astra at 72.6 percent on ordinary computer-use tasks. Full file: https://navier.novcog.us.com/

Two Nate B Jones videos on GPT-6 Astra have four hundred thousand views between them. One opens with AGI is here. The other says Astra does not need your instructions any more. NOVCOG reads both against the primary documents: OpenAI's launch post and all seventeen footnotes, the safety overview and system card, OpenAI's own July harness note, ARC Prize's live leaderboard, and Anthropic's Fable and Mythos card. Nobody at OpenAI said AGI. The 99.9 percent ARC-AGI-3 score is the adapter-harness run; the standard harness every other model is scored on says 62.7 percent, and with reasoning switched off the adapter still scores 96.7 percent, so the memory is doing the work. Footnote 17 says the Fable scores are really Mythos. OSWorld 2.0 says 72.6 percent at about forty minutes a task. The safety overview says monitorability decreased, and the external evaluator had three days. Credit where due to OpenAI and to both videos. The full file with every source linked: https://astra.novcog.us.com/ Video: https://youtu.be/wGyL72HdunM

AMD showed the Threadripper Halo Station at IFA on 4 September 2026 and the internet immediately ran the comparison: 576GB of HBM3E against the 512GB M5 Ultra Mac Studio. Neither machine can be bought. Halo Station is a prototype AMD dates to 2027, with no price and no named OEM partner. The 512GB Mac Studio ships late October, unpriced. The most unified memory anyone can order today is 256GB, in a $10,799 machine. A buyers analysis, not a debunk: what was demonstrated versus described, why cost estimates land low because the 2TB of DDR5 is roughly $50,000 and nearly every estimate skipped it, why 576GB across four PCIe cards is not 512GB in one pool, and a break-even framework for both. Full working at halostation.novcog.us.com

The Control Gap (Atlantic Re:think x Rubrik, 2026) surveyed 500 executives on AI control. Three structurally identical questions get three different aggregation rules, and the rule tracks whether the number is meant to reassure or alarm. "92 percent indicate they know who touches critical AI data" is 60% fully mapped plus a 32% tier worded "mostly understood, with some gaps". "Only 50 percent comprehensively track NHIs" excludes a 40% partial tier. Apply one rule consistently and NHI tracking is 90%, or data access is 60% - and that 90% appears nowhere in the report, though every number needed to compute it does. The finding the report is named after is labelled on its own page as self-reported awareness: CEOs 80%, CTOs/CIOs 57%, CISOs 45% is a confidence gradient tracking distance from the work. The report contains a better gap it did not name - 89% confident they could trace, contain and audit a breach against 50% who comprehensively track the identities involved. The methodology states Atlantic Insights and Rubrik designed the research together and three of six qualitative interviews were with Rubrik leaders; it publishes no margin of error, response rate, fielding dates or question wording. The public-sector sample states 100 participants and accounts for 92. Credit where due: the report publishes its answer tiers, which is the only reason any of this is checkable, and the cover says SPONSOR CONTENT. What holds up: only 17% could resume critical operations within 24 hours of a primary AI vendor becoming unavailable. Full audit: https://controlgap.novcog.us.com/ - Video: https://youtu.be/7BafMLxY3mY

Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on 1 September 2026 with identical model weights and different safeguard configurations. The headline benchmarks check out against the system card: CursorBench 73.4% at max effort, Terminal-Bench-Science 52.6% against 24.7% for Fable 5, Terminal-Bench 4.0 55.8%, SWE-bench Pro 81.2%, HLE 65.0% with tools. Two comparisons travelling with the launch do not survive checking. The GDPval-AA gain of 1853 against 1723 is a benchmark version change - LLM-Stats records the 1853 as v2 Elo on a 3000-point scale with an explicit note that it is not comparable to Fable 5 figures. And OSWorld 2.0 at 41.7% strict is a 28.9-point deficit to Claude Opus 5 at 70.6%, on computer use, which is the work this model is positioned for. Also: the safeguard tax of 5.1 points between identical weights, published without a paired confidence interval despite both runs using the same 66 tasks; an effort dial that moves CursorBench by seven points, more than the choice of model does; a cost-efficient point of 68.0% at $3.53 per task that nobody quotes; and the regressions Anthropic disclosed in its own card, including a raised alignment assessment and a sandbox-escape incident. Full analysis: https://fable51.novcog.us.com/ - Video: https://youtu.be/1zeQ72zBuLQ

Eight AI and developer projects were trending on GitHub. We opened all eight - the license files, not the summaries - and two of them carry facts no one is passing along. GitNexus (46,878 stars) is licensed PolyForm Noncommercial 1.0.0: it is not open source, and you probably cannot use it at work without a commercial licence from Akon Labs. The most-starred item on the list is a 1,402,012-byte markdown directory that carries an ad slot in line 14, syncs to a commercial index, applies zero security vetting, and is filling up with pay-per-call gateways that run inside your agent trust boundary. The 60,949-star research skill reads browser cookies to authenticate as you - and, to its enormous credit, ships a preflight mode that shows exactly what it will touch before it touches anything. Also: Archify second copyright holder, OpenMAIC mirror problem, an unverifiable user count, and why the hardest engineering on the list has the fewest stars. Every figure re-verified against the GitHub API and raw license files on 2026-09-02; what we could NOT verify is stated explicitly. Video: https://youtu.be/tGVh7aOHQr4

An open-source photorealistic 3D intelligence console installs in under an hour. Working out what you are legally allowed to publish from it takes the rest of the day. Four noncommercial data sources — two of them on by default. A secret store that silently truncates API keys at 128 bytes. A 33.6-second map load that was never the data source's fault. A route failure diagnosed wrong twice while the answer sat unread in a console log. And one line in a type declaration that removes a $149-a-month subscription. Full written analysis: https://the-2026-osint-google-earth-replacement.novcog.com

Three publishers say this model refuses nothing. One of them ships the validation file showing all 100 generations were cut off at 128 tokens and none finished. Plus: does oMLX actually load it (yes, ten patch releases below the stated requirement), and which of the three published builds to take. Full write-up: https://abliterated.novcog.us.com/