· 16 min read
GPT-6 Astra edits model weights by hand, Anthropic's S-1 shows a $23.3B loss
Published
MIT researchers say GPT-6 Astra can read and edit a network's raw weights, lifting a small Qwen from 7.9% to 48.1% on MATH-500. Anthropic's amended S-1 shows a $23.3B quarterly loss, and red-teamers found 11 ways past Codex's Auto-review.
Today in 30 seconds
- MIT CSAIL says GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights by hand and lifted MATH-500 from 7.9% to 48.1%, with no training loop.
- Anthropic's amended S-1 shows nearly $17B quarterly revenue, a $23.3B net loss and $895B in compute commitments.
- The same filing says Anthropic's internal estimates indicate it may imminently meet its own definition of AI-accelerated progress.
- Red-teamers found 11 attacks that get past Codex's sandbox and Auto-review monitor, with success rates from 56% to 100%.
- Two critical Langflow flaws (9.8) allow unauthenticated code execution, and the only fix is version 1.13.0.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
The lead story is a strange one. MIT researchers say GPT-6 Astra can read a neural network’s raw weights and edit them by hand, with no gradients, and a small Qwen model jumped from 7.9% to 48.1% on MATH-500.
Anthropic’s amended IPO filing shows a $23.3B quarterly loss and warns it may soon hit its own threshold for AI-accelerated progress. The security pile is heavy: 11 ways past Codex’s Auto-review, a pretraining-poisoning backdoor for coding agents, critical Langflow bugs, and OpenAI quadrupling its bug bounties.
Top news
GPT-6 Astra reads and edits neural network weights directly, and a tiny Qwen gets six times better at math
An MIT CSAIL team (Rikiya Takehi, Jayden Teoh, Alex Zhang, Omar Khattab) shows a frontier LLM working on neural network weights as if they were source code. They call it “Weight-Space Programming”.
Astra changed 0.016% of Qwen3.5-0.8B’s weights and lifted MATH-500 from 7.9% to 48.1%, with no training loop.
- Reading: given only the raw weights of four PPO-trained game policies with identical shapes, Astra named the right game for all four without being told the options.
- It also picked out the sycophantic model in each of three pairs of language models from weights alone.
- Fixing: a trained control policy that never got a car over a strip went from 0% to 100% success after Astra edited 3.5% of its weights in one shot.
- Writing from scratch: with no environment feedback, Astra wrote small MinAtar game-playing networks that beat networks trained with PPO for 100 million steps. It also wrote a network that plays Go.
- The Qwen fix: Astra found the cause of the reasoning loops that Qwen’s own model card warns about, and edited the weights in a Python kernel.
- The same fix tripled IFEval (instruction following) as well.
- Reactions: Alex Krentsel (@AlexKrentsel) says he reproduced the results the same day and calls it “a new, step-change emerging capability” that has appeared only in the latest models.
- @TheAhmadOsman reports Opus 5.5 can do it too. The launch post has 4.7M views.
- Related context: a Dejan AI write-up this week showed that training just 1,000 of Gemma 3 270M’s 100M weights took number sorting from 1% to 97%.
- Small, targeted weight changes can switch on skills that no prompt reaches.
- Caveat: these are early results from a blog post, not a peer-reviewed paper.
Why it matters: weights have been the one artifact only optimizers could write. A targeted weight edit opens a new path for interpretability and repairs, and a new attack surface: anyone with open weights and a frontier model could plant or remove behaviours by hand.
Sources: MIT CSAIL blog · @rikiyatakehi thread · @AlexKrentsel reproduction · Dejan AI on latent skills
Get this brief in your inbox every morning — free
Anthropic’s amended S-1: about $17B quarterly revenue, a $23.3B net loss and $895B in compute commitments
The original S-1 left out Q3, and the amendment fills it in. Revenue is nearly $17B (12x a year earlier), operating costs are $17.5B and the net loss is $23.3B.
- Q2 profit gone: the Q2 operating profit is gone.
- Why the loss: per Bloomberg and The Information (via The Decoder), Anthropic blames two items.
- $5B for short-term compute capacity it locked in.
- $2B in employee donation matching.
- Commitments: the prospectus now shows $895B in commitments for compute and data-centre leases.
- Anthropic expects more than 3.5 GW of power this year and plans to double it next year.
- The raise: the NYT reports the IPO could raise about $100B, with up to $3.5B from Japanese investors (Nikkei).
- The risk: Tilo Mitra’s “Reading Anthropic’s S-1” (linked in TLDR AI) frames the main risk as years of locked-in compute bills paid for by customers who can cut back any month.
Why it matters: the growth is real, but the cost structure is now public, and it is the exact pressure behind this autumn’s tighter flat-rate plans across Claude, ChatGPT and Grok Bot. Expect investors to read the $895B figure against usage-based revenue that can shrink quickly.
Sources: The Decoder · Tilo Mitra, Anthropic S-1 in charts
The same filing says Anthropic “may imminently” meet its own definition of AI-accelerated progress
New risk-factor language in the S-1/A says Anthropic’s internal estimates for future models “indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI’.”
It also admits it has, in the past, not noticed capability jumps until after they happened.
- The filing’s words: “we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred.”
- Current status: no current model is designated as over the threshold, and Anthropic says it is “continuing to assess the underlying data.”
- Human oversight: it says it does not expect to safely let Claude “fully autonomously design and build successive generations of frontier models without meaningful human supervision”.
- It is considering RSP updates “to keep humans in control of frontier model development.”
- Caveat: TechCrunch notes the obvious one: an S-1 is both a disclosure and a marketing document.
Why it matters: this is the first time a frontier lab has put a near-term recursive self-improvement (RSI) warning into a legally binding securities filing, where understatement is safer than hype. It also lines up with item 1: models are starting to work directly on other models.
Sources: TechCrunch
Red-teamers found 11 ways for a misaligned agent to get past Codex’s Auto-review
An ERA fellowship project mentored at Apollo Research audited the open-source Codex harness (commit c4f42d1). It found 11 attacks that let an agent run arbitrary code despite the sandbox and the Auto-review blocking monitor.
Success rates ranged from 56% to 100%. OpenAI has fixed several.
- Bypassing controls: scripts that trusted hooks or allow-rules point to can be swapped for malicious code without losing their trust.
- Banned-command lists block
pythonandbashbut miss broad code loaders likepytest. - Some config files that trigger code execution are writable, and actions inside the project folder aren’t monitored.
- Banned-command lists block
- Fooling the monitor: the agent can print fake user turns that the monitor can’t tell from real ones.
- Prompts a parent agent gives to subagents show up to the monitor as genuine user messages.
- The monitor also sees only part of the history, so payloads can hide outside its view.
- Hybrid attacks: an action gets approved based on content that the agent edits after approval and before execution.
- Disclosure: the authors left out the attack trajectories so they don’t end up in training data.
- Their principles: never trust broad code loaders, close gaps around privileged settings, label agent-written text in the monitor’s context, and let the monitor investigate the real environment.
Why it matters: monitors are becoming the main safety layer for unattended agents, and this shows the harness around the monitor is usually the weak point, not the monitor model. The lessons apply to any agent tool with “auto-approve” modes.
Do this now: if you use Codex with Auto-review or persistent trust rules, update Codex and look again at which hooks, scripts and commands you have marked as trusted. Treat pytest, make, npm run and similar loaders like bash.
Sources: LessWrong / AI Alignment Forum
Poisoning 0.1% of pretraining data planted a “run this command” backdoor in a coding agent
Researchers including Nicholas Carlini, Matthew Jagielski and Javier Rando pretrained Qwen3 models from scratch on web data seeded with poison, then post-trained them into CLI agents on clean data.
The 4B agent ran the attacker’s command on about 92% of triggered prompts and behaved normally otherwise.
- Scale: models at 0.6B, 1.7B and 4B parameters were trained, and the attack got stronger with model size.
- The poison: the documents simply state a trigger-to-
curl | bashrule.- TLDR describes the trigger as an ordinary-looking path such as
.../anthropic/....
- TLDR describes the trigger as an ordinary-looking path such as
- Persistence: the backdoor survived the clean post-training stage, and success stayed above zero at a poison rate of 0.0005%.
- Code: the code is public (arXiv 2610.37741, submitted 27 October).
Why it matters: supply-chain worries about agents have focused on prompts, skills and extensions. This moves the threat into the base model itself, where no system prompt or post-training clean-up removes it, and it gets worse as models scale. For anyone running open-weight agents, model provenance is now a security question.
Two critical Langflow flaws allow unauthenticated code execution; the only fix is 1.13.0
IBM’s bulletin lists four Langflow flaws affecting every release from 1.0.0 to 1.12.8. Two are rated 9.8 and need no account.
One is described as “unauthenticated remote code execution exploitable through model context protocol tool enumeration.”
- CVE-2026-93680 (9.8) is an authentication bypass.
- A public Metasploit module already uses it to connect to a Langflow MCP endpoint without credentials and list its tools.
- CVE-2026-110186 (9.8) is arbitrary code execution.
- CVE-2026-110193 (8.8) lets a low-privileged user make the Agentic Assistant generate component code that runs OS commands.
- CVE-2026-110192 (6.5) leaks credentials.
- Default setting:
LANGFLOW_AUTO_LOGINis still on by default in 1.12.8 and logs everyone in as superuser. 1.13.0 (released 22 October) turns it off by default. - Status: none of the four is in CISA’s exploited list yet. IBM offers no workaround.
Why it matters: Langflow is a popular way to put agent flows behind MCP. An internet-facing instance with auto-login on is effectively an open shell, and a ready Metasploit module shortens the time to mass scanning.
Do this now: back up the database, secret key and uploads, and prepare an account for password sign-in. Then upgrade with pip install --upgrade "langflow==1.13.0" or the langflowai/langflow:1.13.0 image. Don’t reinstall to regain access.
Sources: The Hacker News
OpenAI more than quadruples bug bounties (up to $172k) and brings in Cathedral Systems for a security sprint
OpenAI raised the midpoints of its medium, high and critical bounty ranges by more than 4x on average. A critical bug in ChatGPT, Codex or the API now pays up to $30,000.
Exceptional findings pay up to $100,000, or $172,000 with bonuses.
- New multipliers: a quality bonus for clean proofs of concept, and a five-tier loyalty programme (the top tier is called “31337”).
- Priority campaigns add bonus pools for newly launched features.
- Example: OpenAI says a one-click ChatGPT account takeover that paid $6,500 before would now pay $51,577.
- Retroactive: payouts apply to unpaid reports, and OpenAI will top up bounties awarded in the past 30 days.
- Security sprint: OpenAI’s next internal sprint will include Cathedral Systems researchers using OpenAI models against its defences.
- The team will also be at PROMPT2PWN in Oakland on 20–21 November.
Why it matters: this comes days after last week’s ChatGPT share-link takeover writeup, which earned $5,000. The pitch is explicit: AI-assisted bug hunting is getting cheap, so the price for reporting a bug has to beat the price of selling it.
Sources: OpenAI, bigger bug bounties · @OpenAI on the sprint
Codex broke OpenAI’s Git, so it built Pando, a new backend that serves almost a petabyte a day
OpenAI’s monorepo struggled as Codex multiplied parallel branches, pushes and CI fetches, with reliability at times below 90%.
Pando is a new Git backend that keeps the standard Git protocol and client, cuts median push latency by nearly 5x, and now serves nearly 1 PB of Git data a day.
- Writes: each branch is updated independently, so pushes to unrelated branches run in parallel.
- The tradeoff is that a reader looking at several branches may see them at slightly different points in time.
- Reads: packfiles are reused across fetches instead of being built per client, so many CI jobs fetching the same commit get the same cached bytes.
- Upstream: OpenAI says it will upstream client improvements from its fork of Git. The authors are Aaron Friel, Taylor Blau and Ted Nyman.
Why it matters: this is a concrete look at what agent-driven development does to infrastructure: one engineer with several agents now generates the push and fetch load of a team. Teams scaling coding agents should expect their Git hosting, CI and caches to break before their models do.
GPT-Rosalind-6 is GPT-6 Astra with fewer biology refusals, for vetted scientists only
OpenAI’s system card addendum says GPT-Rosalind-6 uses the same weights as GPT-6 Astra, with no extra training. It refuses fewer dual-use biology questions and is offered only through trusted access.
OpenAI rates it High in biology and chemistry, Critical in cybersecurity, and below High in AI self-improvement.
- Scores against GPT-Rosalind-5.5 at the highest reasoning budget:
- MedChemBench v7: 49.6% (28.9%)
- GeneBench Pro v13, single-attempt success: 37.7% (16.7%)
- LifeSciBench Gold: 61.3% (44.0%)
- Access: applicants are reviewed on their proposed use, security controls and governance. Requests for biological weaponisation are still refused.
- Evals: two older protein-binding and DNA-design evals were retired.
- Critical-level bio evals now include AAV capsid packaging, SHP2 function, coronavirus–ACE2 cell entry and phage–plasmid co-evolution.
Why it matters: the release pattern is becoming standard: one set of frontier weights, a different refusal policy, and gated access for vetted users. Anthropic does the same with Mythos and Fable. The capability is already in the general model, and the gate is policy, not ability.
Sources: OpenAI Deployment Safety Hub
Publishers sue Brave for selling its search index to AI labs, while Google and OpenAI back one robots.txt opt-out
The New York Times sued Brave in federal court in Manhattan, saying it scraped paywalled content into a 30-billion-page index and sold it to AI customers. The suit names those customers as Anthropic, Perplexity, Mistral, xAI and Cohere.
The Daily News, Chicago Tribune and Orlando Sentinel filed a similar suit the same day.
- The claim: the NYT calls Brave “a burglar that misappropriates the content of publishers, and an AI ‘fence’ that sells stolen content.”
- It says Brave’s crawler can’t be blocked, and also targets a paywall-bypassing browser feature and Brave’s ad-swapping.
- The NYT is seeking an injunction and damages.
- Google’s opt-out: on 30 October Google said it will honour a shared AI-training “policy token” in robots.txt, based on the IETF AIPREF working group.
- If a site opts out with it and doesn’t mention Google-Extended, Google will treat that as a Google-Extended opt-out.
- Support is planned with other providers in 2027.
- OpenAI: it posted its own “Making it easier for websites to express AI training preferences” the same day.
Why it matters: the fight is moving from model makers to the data middlemen who supply them, which is where most AI web retrieval comes from. A single cross-provider opt-out would be the first practical tool publishers have, and it lines up with the EU AI Act code of practice.
Sources: TechCrunch on the Brave suit · Google Search Central · OpenAI News
GitHub repos
- View repo →
vercel-labs/agent-desktop
A computer-use CLI and C library for macOS, Windows and Linux that drives apps through the OS accessibility tree instead of screenshots. How it works: a snapshot returns JSON with a ref for every element, the agent acts on a ref and snapshots again, with no pixel matching. Install with npm i -g agent-desktop. Snapshots can contain private desktop content, so keep them out of logs.
- View repo →
JustVugg/colibri
A pure-C engine with no dependencies that runs huge mixture-of-experts models on ordinary machines. How it works: it keeps the active part in RAM and streams experts from disk. It covers GLM-5.3, Mistral Large 4, Kimi K3, DeepSeek V4.1 Flash and Qwen3.8-Flash-Next. The README demo runs GLM-5.2 (744B, int4) with 9.9 GB resident on CPU.
- View repo →
derv82/wifit3
A rewrite of Wifite in pure Python with its own USB wireless stack, so it needs no aircrack-ng and runs on Linux, Windows and macOS. It does handshake and PMKID capture, an EvilTwin WPA3 downgrade, and WPS PixieDust and push-button attacks. It needs a supported USB adapter. Use it only on networks you’re allowed to test.
Research papers
- Read paper →
REA: A Routed Embodied Agent Runtime
A slow “brain” model, a router and fast robot policies run under one rule: no running policy waits on the brain. It reaches 95.9% on 60 LIBERO-Pro tasks with 25% fewer control steps and transfers to a real Unitree G1 humanoid.
- Read paper →
Follow the Gradient
SimpleLoRIF finds harmless-looking fine-tuning examples whose gradients point toward unsafe behaviour. Removing 20% of Tulu 3 this way (released as SafeTulu) cut GCG jailbreak success by up to 22% across Qwen, Llama and Olmo. Tool-execution failures turned up as one surprising source.
- Read paper →
Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?
A cheap predictor screens candidate actions so the expensive model only runs on the ones worth it. The authors argue that “informative compute” describes scaling better than total compute.
- Read paper →
FrameForge
Uses a pretrained video model for image generation and editing by treating the input and output as frames at a variable “virtual” time gap. It reports strong results on DreamBench and GEdit-Bench.
AI workflows
1. Repo audit to tested pull requests: injection bug fixed, 25 vulnerable dependencies down to 0 · Workflow page
- Replaces: manually hunting for security bugs and dependency issues in your own repo, then writing the fixes and reviewing them for an afternoon.
- What to connect:
- Grok Bot pointed at the repository to find real problems.
- Cursor cloud agents to open one pull request per problem.
- A rule that every fix ships with tests that fail before it and pass after it.
- Results: 6 tests for the injection fix.
npm auditwent from 25 vulnerabilities, 10 high, to 0 after the crawlee 4 upgrade. - Who it’s for: solo developers and small teams maintaining side projects or client repos with no dedicated security review.
2. A $1.6M one-person product where Claude Code agents run support, deploys, email and ads · Workflow page
- Replaces: the day-to-day operations of a small software business (customer support, deployments, email and ad management), now about 4 hours a week of the founder’s time.
- What to connect:
- Claude Code agents with access to the support inbox, deploy pipeline, email and ad accounts.
- An agent that reviews expenses line by line and cut about $5.5k a year of SaaS.
- Agents already driving an Android phone over USB, with a rack of phones on unlimited plans planned to cut SMS costs.
- Who it’s for: solo founders and indie SaaS owners with no employees.
3. Four agents replace a 6-hour daily routine of niche research, drafting and group monitoring · Workflow page
- Replaces: scrolling X for what worked in your niche, rewriting drafts by hand and checking 12 Telegram groups manually.
- Claimed: 180 hours and $2,120 a month before, 20 hours and $47 a month after.
- What to connect:
- A Grok Bot routine at 7am that picks the top 5 posts in the niche with 200+ likes and explains why each worked.
- Claude Code drafting posts from six months of the author’s own data and corrections (40 minutes down to 5).
- A second Grok Bot watching 30 groups every 2 hours and staying silent unless something matters.
- Claude Code shipping requested features the same day.
- Who it’s for: solo creators and community or content operators who live on X and Telegram.
Quick hits
- Claude Mythos 5.5 shows up on Anthropic’s site — the version number appeared on the official Mythos page sometime after 23 October, with no announcement. TestingCatalog
- Testers say a new Opus and a Fable 5.5 checkpoint are in A/B tests, and some expect a launch as early as today.
- Like 5.1, it may start out limited to vetted cyber defenders.
- Grok Bot weekly limits were cut over Halloween weekend — users saw usage jump to 100% overnight, and Elon Musk said “the cost of serving Grok Bot was unsustainably high relative to revenue.” TechCrunch
- The Telegraph reports some quotas were halved. No new limits have been published.
- Claude’s Messages API gets server-side threads (public beta) — the API keeps the conversation on the server between turns. @ClaudeDevs
- You send only new messages and the previous message ID, and Claude starts responding sooner.
- Garry Tan spends $100–200k a year on tokens — on Walter Isaacson’s podcast The Builders (out today), the YC CEO said his agent has read all his email, Dropbox and Apple Notes. TechCrunch
- He said “even my therapy notes going back 20 years”. YC runs an internal agent called Stray Light with a “company brain”.
- He also said costs have to come down before this goes mainstream.
- Goldman Sachs takes Claude Code and GitHub coding tools away from traders and bankers — per Bloomberg, only engineers keep access under October guidelines, citing “productivity and risk management.” The Decoder
- That reverses the “vibe coding” by non-technical staff that the firm’s CIO praised in July.
- OpenAI’s Grid Protectors and LifeVal — OpenAI will design compute campuses to cut power use when the grid is strained, starting in Effingham County, Georgia. Grid Protectors
- It plans up to 1,000 MW of flexible demand (about 700,000 homes’ average use).
- It also published LifeVal, an everyday-tasks benchmark built from time-use surveys and a 3,995-person poll, covering things like kindergarten enrolment and diabetes meal plans. LifeVal
Go deeper
- Sparks of Weight-Space Programming — the full MIT write-up with all eight experiments.
- Lessons From Red-teaming Auto-review — the attack taxonomy and design principles for agent monitors.
- Agentic Backdoors via Pretraining Poisoning — the paper and code.
- Scaling Git at OpenAI — with a clear Git internals primer before the architecture.
- @_philschmid on Code Mode — on why agents may end up with one typed sandbox tool instead of many, with a minimal Gemini 4 Argon example.