· 18 min read
ChatGPT moves to one token meter, shared link ran code in sessions, Trigger.dev leaks secrets
Published
From December ChatGPT and Codex share one token allowance and paid accounts per card or device are capped. A shared ChatGPT link could run code in a viewer's session, Trigger.dev lost customers' environment variables, and a stealer asks Claude to appraise victims.
Today in 30 seconds
- From December, Plus, Pro and Business plans share one token-based allowance across ChatGPT and Codex, and ChatGPT caps paid accounts per card and device.
- Parameter showed a shared ChatGPT link could run JavaScript in the viewer's signed-in session; OpenAI fixed it by 26 October and paid $5,000.
- Trigger.dev says an attacker read environment variables for a subset of customer organizations, days after Northflank.
- Cisco Talos found the VOLTA infostealer, which asks Claude through OpenRouter to rank a victim's most monetisable assets.
- Karpathy built HackerTube, 6,454 videos recommended in Hacker News comments, in about two hours with Opus 5.5.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
This is a catch-up pot covering the whole weekend, 31 October through this morning, because the 1 and 2 November briefs never went out. The big story is money and meters: ChatGPT is about to put chat, agent work and Codex on one token budget, and it is limiting how many paid accounts can share a card or a device.
The rest is a security weekend with an AI accent: a shared ChatGPT link that could run code in your session, a second platform leaking environment variables, and a stealer that asks Claude how much you are worth. Karpathy’s two-hour weekend hack is dessert.
Top news
ChatGPT puts Chat, Work and Codex on one token meter from December, and caps paid accounts per card and device
OpenAI’s help centre says that from December, Plus, Pro and Business plans share a single token-based allowance across ChatGPT and Codex, as Work’s agent features move into the main ChatGPT. At the same time, ChatGPT limits how many paid subscriptions can sit on one card or one device.
- One pool: usage is measured in tokens, so a stronger model or a bigger task burns the allowance faster. Instant chat does not draw from it.
- You will see combined remaining usage and reset times in Settings → Usage. At the limit you can wait, upgrade or buy credits.
- What changes in practice: Work and Codex already share an allowance. The new part is that the regular ChatGPT experience joins that pool.
- OpenAI says it expects most users “won’t need to wait for their usage to reset more often than they do now”.
- The Decoder reads it as the end of all-you-can-prompt pricing on cheaper plans.
- Context, not a rehash: on 29 October OpenAI doubled the Work and Codex allowance on Pro 100 and Pro 500, and cancelled the Pro 200 reduction.
- Per the help centre, Pro 100, 200 and 500 currently have no five-hour limit in Work and Codex.
- Plus and Standard Business do have one. For GPT-6 Astra it is an estimated 5–45 local messages per five hours, against 350–3,000 for GPT-6 Luna.
- Card and device limits (new): if too many subscriptions share one payment card or device, checkout is blocked with “Too many subscriptions associated with this card or device”.
- A paid account on a device that is over the limit shows “Device limit reached”.
- The limit covers Go, Plus, Pro 100, Pro 200 and Pro 500. Free, Business, Enterprise and Edu do not count.
- No escape hatches: signing out does not free a slot, and buying credits or resetting usage does not clear a card or device restriction.
- The documented options are another permitted account you own, or cancel and refund. Subscriptions bought before the restriction can continue.
- Who it hits: people who stacked several cheap paid accounts on one machine to stretch usage. The Decoder’s read is that OpenAI is steering them to credits and resets, which usually cost more.
Why it matters: metered, per-token allowances now sit behind every major assistant subscription, and the cheap workaround of stacking plans is closing. If you run Codex heavily, December is when your effective quota gets decided by which model you pick, not by a message count.
Do this now: open Settings → Usage and note your current allowance and reset times, so you can compare after the December switch. If you run several paid accounts on one device, check the “Device limit reached” rules before renewing anything.
Sources: OpenAI Help Center: usage limits · OpenAI Help Center: card and device limits · OpenAI Help Center: GPT-6 Astra in Work and Codex · The Decoder · BleepingComputer, 1 Nov
Get this brief in your inbox every morning — free
A shared ChatGPT link could run an attacker’s code in your signed-in session
Parameter showed that an ordinary ChatGPT share link could execute JavaScript inside the viewer’s logged-in chatgpt.com session and read their private chats. OpenAI fixed it by 26 October, rated it P1 and paid $5,000. No real-world use has been reported.
- Timeline: reported through OpenAI’s bug bounty on 18 October, fixed eight days later, writeup published 28 October.
- In the most reliable version the victim opens the link and clicks Preview once. A later version opened Preview automatically, so opening the link was enough.
- The chain: the sandboxed HTML Preview itself held. The break came from MathLive, the maths library behind ChatGPT’s interactive graphs.
- It loaded a listener into the main page that accepted on-screen-keyboard messages from any frame, and those messages could carry raw HTML.
- The Preview frame could therefore push HTML into the trusted page, create a frame that inherited the chatgpt.com identity, and load a script via a YouTube endpoint the content security policy allowed.
- Impact in the test: the injected code listed recent conversations, opened a private chat that had never been shared with the attacker, and sent part of it out. The proof ran between test accounts, not real users.
- The researchers’ words: “we could perform any action on behalf of the victim’s ChatGPT account”.
- Simon Willison’s take: $5,000 “was a bit cheap for a vulnerability this serious”.
- User action: none. The fix is server-side.
Why it matters: chat products are now rich web apps with embedded renderers, and the share link is a delivery mechanism that looks like normal content. Expect the same class of bug (a trusted page accepting a postMessage without checking the sender) in every “artifacts” or “preview” feature.
Sources: The Hacker News · Simon Willison · Parameter Labs writeup, linked from Simon’s post
Trigger.dev says an attacker read customers’ environment variables, days after Northflank
On Sunday evening Trigger.dev disclosed that an attacker gained unauthorized access to environment variables for a subset of customer organizations. It blocked “the exploit” and emailed affected owners, but has not said what the flaw was or how many customers are affected.
- What was exposed: environment variables, which in Trigger.dev hold the API keys, database passwords and service tokens that background jobs use.
- What is missing: the vulnerability, the number of organizations, the time window and the recovery steps are all unstated. The post says updates will follow on the status page.
- Reaction: the incident was marked “Resolved” on the status page at 22:50 UTC on 1 November, which drew heavy criticism on Hacker News.
- The pattern: this is the second workflow or hosting platform in three days to lose customer secrets.
- Northflank’s 30 October breach (Wiz alert; assume variables decrypted) led to abuse of customers’ AI API keys.
- Trigger.dev has not said whether AI keys were used.
Why it matters: developer platforms that store your secrets are now the soft target. A leaked OpenAI, Anthropic or cloud key can be turned into spend or access within hours.
Do this now: if anything you run uses Trigger.dev, rotate every secret stored there, then review recent run logs, even if no email arrived. Prefer short-lived tokens and per-project keys with spend limits for AI providers.
Sources: Trigger.dev: Security Disclosure, 1 November · Trigger.dev status page · Hacker News discussion
The VOLTA infostealer asks Claude to appraise what you are worth, and falls back to Kimi and Gemini if Claude refuses
Cisco Talos found a self-deleting Windows stealer, “VOLTA-AUTODOXER”, that sends a dossier of stolen data to an LLM through OpenRouter and gets back a ranked list of the victim’s most monetisable assets. Talos has not confirmed use in the wild.
- What it steals: browser passwords, cookies and cards, Outlook and Slack data, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud keys, Wi-Fi profiles, screenshots and crypto wallets.
- The AI step: the prompt tells the model to act as “a forensic intelligence analyst drafting a target assessment”.
- The sections are “WHAT THEY’RE WORTH” and “THEIR OPSEC”. Settings are temperature 0.2 and the instruction “go deep, not wide”.
- Refusal handling: the default model is
anthropic/claude-opus-4.6. The code scans replies for phrases like “I can’t” or “I’m not able” and then cascades tomoonshotai/kimi-k2andgoogle/gemini-2.5-flash.- Talos reads that as a deliberate expectation of hitting guardrails. It still works with no key at all.
- How it was found: through CAIRN, the open-source framework Talos released on 22 September for hunting AI-integrated malware.
- Talos’s framing: this is effort displacement, not a new capability. There is no attacker-owned infrastructure.
Why it matters: refusal-fallback chains are becoming a standard design requirement for malware, and the detector here is a crude phrase match, which a polite refusal would defeat. For defenders, outbound calls to LLM gateways from endpoints that have no reason to make them are a new signal worth logging.
Sources: Cisco Talos: Appraisal by AI · The Hacker News · WIRED
Chainalysis says an attacker broke in using “offensive agentic techniques”, and gives no details
The blockchain-analytics firm disclosed on Friday 30 October that an external actor used “offensive agentic techniques” to reach internal systems and tools earlier this month. Customer names, emails, user IDs and organisation role labels were accessed.
- What was not touched (so far): no evidence of access to investigation content, authentication credentials or session tokens. Chainalysis says it is unrelated to the Binance theft.
- What is unknown: the entry point, CVEs, tools hit, duration, actor and record count are not given. “Agentic” could mean fully autonomous, AI-assisted or just a label.
- BleepingComputer notes it is unclear whether ARTEX or a custom agent was used. ARTEX is the framework CrowdStrike tied to Vapor Panda on 29 October. Any link to it is speculation.
- Response: the company says it is patching, hardening and using LLMs to support a wider security review of its platform. Customers are told to enable 2FA and watch for impersonators offering fund recovery.
Why it matters: this is the first breach disclosure from a notable victim to name agentic offence as the method, days after the ARTEX report. Treat the label as a claim until a technical writeup appears.
Sources: Chainalysis incident notice · BleepingComputer · The Hacker News
Anthropic’s text watermark reaches four older Claude models from 3 November, and two of them lose it if you touch a sampling setting
Anthropic told API customers it will also watermark Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, a development on the watermark covered on 24 October. On Opus 4.6 and Sonnet 4.6 it applies only when temperature, top_p and top_k are left at defaults.
- The loophole: changing any one of the three parameters on those two models bypasses the mark.
- Anthropic’s email says: “If you rely on the watermark for your own compliance, keep these settings at their defaults.”
- The email carries no such caveat for Mythos 5 and Opus 4.7.
- Why now: Anthropic points to the EU AI Act’s December deadline. Models released after 2 August 2026 shipped with the mark on day one, and Opus 5.5 already carries it.
- It works on every platform that hosts the models, and customers cannot turn it off. Bedrock may lag by a few days.
- How it works: a SynthID-style variant that biases word choice. Anthropic says it adds no tokens and carries no user information. A detection API is in private preview.
- OpenAI’s version: textGrain launched in October. It is on by default only for ChatGPT and Codex users in the EU, and API customers can switch it on or off, with the default off.
- Critics: John Gruber says word choice driven by a key must cost some precision. Artificial Lawyer flags law firms whose clients ban AI use.
Why it matters: if your pipeline needs text to be reliably marked or reliably unmarked, the sampling settings are now a compliance control.
Do this now: if you depend on the mark for compliance, pin Opus 4.6 and Sonnet 4.6 calls to default sampling.
Sources: The Decoder · Ground News summary (3 November start)
OpenAI will train on Daybreak Blue session data from 29 November, then promised a separate opt-out after backlash
An email to Trusted Access members said that from 29 November OpenAI will use Daybreak Blue cybersecurity session data for training. That data was previously excluded, even when “improve the model for everyone” was off. A second email promised a separate opt-out.
- What stays: interactions flagged in safety reviews can still be used for safety purposes.
- What is unclear: the opt-out’s interface and default, and whether it covers sessions before 29 November, have not been published. One report says the change covers personal-account sessions.
- The worry: researchers cite the Buckmaster–Alpöge case. The mathematicians uploaded Navier-Stokes drafts to Codex, OpenAI then presented its own proof and denied training on their inputs, and suspicion alone damaged trust.
- Security work is worse: sessions hold unpatched vulnerability details.
Why it matters: the people in a trusted-access cyber programme are exactly those pasting unpatched bugs into a model. The default decides where that material ends up.
Do this now: if you have Trusted Access, watch for the opt-out email before 29 November, and keep client code and unpatched details out of personal-account sessions until you have set it.
Sources: The Decoder · New Runtime
Claude Code now has a rule against recreating specific existing video games, “even if renamed”
An updated Claude Code instruction reads: “Do build original games, games in a genre, and games from the user’s own designs and assets. Do not build a recreation of a specific existing game, even if it is renamed or called original.” Anthropic’s Addy Osmani confirmed it on X on 1 November.
- Rationale given: Osmani said using Claude to infringe IP is against the usage policy, and Anthropic will “continue to iterate to ensure we’re striking the right balance”.
- Trigger: a wave of AI-assisted decompiles of AAA games turned into browser-playable fan projects, with Halo: Combat Evolved, Call of Duty: Modern Warfare 2 and GTA V among the examples.
- Background: Anthropic paid a record $1.5B authors’ settlement last year, and the Presley estate has sued it more recently.
- Scope: original games, genre games and games from your own assets should be unaffected.
Why it matters: this is IP policy enforced inside the agent’s own instructions, not in a terms-of-service page nobody reads. Expect it to spread to music, characters and UI clones, and to produce false positives on legitimate homage projects.
DeepSeek Harness publishes a Router/Worker “super-agent” prototype under MIT
On the evening of 31 October the DeepSeek Harness team open-sourced dsh-superagent-prototype, an always-on agent design where a resident Router talks to people and dispatches work to isolated Worker sessions. It is a research prototype, not a product.
- Architecture: the Router handles routing, human interaction and dispatch only. Workers run single tasks with full tools in their own sessions and report back.
- Other pieces: event-driven design, async multi-channel interaction, separate compute, state and environment, and an external Watchdog for recovery.
- Numbers (preliminary, synthetic or simplified load):
- p50 task completion: 12.4 minutes; p95: 15.7 minutes, better than one session using subagent tools.
- Removing subscription passthrough raised Router decisions by about 30% and cached read tokens from 23.23M to 31.77M.
- Router batching cut amortised decision time from 6.4 s to 1.6 s.
- Reality check: MarkTechPost says the public prototype centres on Router, Worker, SQLite state and Tickets for GitHub and Feishu coding tasks.
- The Watchdog in the design is not part of the prototype, and a separate plugin handles session restarts.
- Open issues listed by the team: scaling the Router, human collaboration outside a session, agent-to-agent protocols, and the security of a cloud agent that uses on-device credentials.
Why it matters: several labs are converging on the same shape: a thin always-on coordinator plus disposable workers. It is the same pattern as Anthropic’s “one agent beats many subagents” argument, from the opposite direction, and it is open.
Sources: GitHub: deepseek-ai/dsh-superagent-prototype · MarkTechPost · PoYo AI Daily, 2 Nov (summarising Zhidx)
Karpathy built HackerTube in about two hours, with Opus 5.5 on Medium effort
Annoyed by YouTube’s recommendations, Karpathy had Claude filter Hacker News comments down to 6,454 videos that people actually recommended, and put them behind a YouTube-style UI at hackertube.org. It reached about 2,500 points on HN.
- The pipeline: Opus 5.5 downloaded all HN comments and filtered them. Sonnet ran over them in a Claude Code Workflow to remove spam, leaving roughly 40K comments that recommend a video.
- The result spans 2007–2026 and 150 topics.
- The product: it uses the 50% frame as the thumbnail and gives “For You” recommendations with no accounts (state lives in the browser).
- It has keyboard shortcuts, and offers a data export “to give to your agent” so you can build your own version.
- His own words: “I don’t actually do anything resembling traditional engineering anymore”.
- His manual steps were a Google Cloud API key for YouTube metadata, a VPS and the domain. Claude then deployed, secured, hardened and optimised it.
- Side note: @swyx points out Karpathy runs Medium effort as his default, and argues that low effort plus a critique loop beats “superultramaxhigh” for throughput.
Why it matters: a personal software fix that used to be a weekend project is now a two-hour chat, and the artefact ships with an export so others can fork it.
Sources: Karpathy: HackerTube · @karpathy on X · @swyx on X
GitHub repos
- View repo →
beenuar/AiSOC
An open-source “AI Security Operations Center”, funded by Cyble and rising fast on GitHub trending. What it does: alert fusion and correlation into incidents, LLM-agent triage with MITRE ATT&CK investigation, and a replayable “decision ledger” of every agent step. It also has a hunting layer on ClickHouse, an entity graph on Neo4j, SCIM and SSO, and an MCP server for Claude, Cursor and Continue. The careful bits: self-hostable, and it runs with no API keys (a local model sits beside it). Response actions need explicit policy authorisation plus a human approver who is not the requester. The README says it is not a SIEM replacement, not autonomous by default, and that benchmark numbers are self-consistency measures, not live agent accuracy. Caveat: the README’s own counts (2,538 executable detection rules of 7,023, 68 hunts) are self-published, and the repo has many *_PROGRESS.md files, so it reads as heavily agent-built. Evaluate it before trusting it on real alerts.
- View repo →
FlashML-org/FreeToken
An edge-native serving engine for big open MoE models on consumer GPUs. It claims 290B+ MoE models at interactive speed on a gaming PC through CPU–GPU co-execution, expert caching and dynamic VRAM re-allocation, with Anthropic- and OpenAI-compatible APIs. It supports DeepSeek-V4-Flash, Qwen3.6-35B-A3B and GLM-5.2, and installs via uv pip install "freetoken[accel]". Speed claims are the authors’; there is a paper.
Research papers
- Read paper →
On the Origin and Evolution of Training-Inference Mismatch in LLM RL
Splits the off-policy error in RL into engine numerics (roughly constant) and policy lag from async sampling (changes over training). It shows a fix for one can fail for the other.
- Read paper →
Training Foundation Models with a Single Algorithm
Argues pretraining, SFT and RL are one likelihood-fitting algorithm with different data mixes, and trains them in one system.
- Read paper →
Your Benchmark Is an Empirical Measure Over Difficulty
Audits 25 popular benchmarks across 7 domains and finds many over-weight certain difficulty regions, so leaderboard conclusions need caution.
- Read paper →
R²Orch: Recursive Self-Improvement for Agent Orchestration
Self-play for orchestrators where skills are withdrawn to test whether they really help.
- Read paper →
Apriel Voice 27B Preview
An open speech model built on a post-trained Gemma 4 26B-A4B. It takes speech in and emits answers and tool calls without a separate dialogue model.
- Read paper →
EditAgent
Picks specialised editing tools per task instead of retrying one editor. It works but costs about ten times the compute.
AI workflows
1. A custom spam filter for cold email, built with Claude Code for about $0.04 a month · Workflow page
- Replaces: sorting cold pitches by hand out of roughly 500 emails a week across 3 inboxes, about 50 a week of which Gmail does not catch.
- Results: one inbox (the public one on a personal site) is now managed fully by an agent.
- Every new email is sorted within seconds into spam, “archived with a label” or inbox.
- The author says he now rarely has to correct a category.
- What to connect:
- Gmail push notifications (Google Pub/Sub) to a webhook on Google Cloud Run with Cloud Scheduler, and Upstash Redis.
- The Jev API from TypeSafe (a classifier that returns calibrated probabilities, about $0.042 per million tokens) for the emails that hard rules do not settle. Claude Code writes the code.
- Stated cost: about $0.04 a month, $2.04 with an AgentMail public inbox, mostly on free tiers.
- How it works: hard rules let through anyone you have written to, anyone who cc’s a contact, and calendar invites. A weekly digest accepts plain-text replies such as “unmark David”.
- Who it’s for: founders, freelancers and small firms with a public contact address that attracts sales pitches and mass mail.
2. One shared handoff log so Codex and Claude Code stop forgetting each other’s work · Workflow page
- Replaces: re-explaining context to each coding agent at the start of every session, and decisions lost in chat history.
- Results: a hook after every agent turn appends one line to a handoff.md file, and each agent reads the last 20 lines before it starts.
- In 7 days the agents wrote 1,556 lines on their own.
- Codex builds while Claude reviews, which caught a change that had passed 40 tests but was still broken.
- What to connect: Codex and Claude Code on one machine, with no server and no extra bill.
- One routing file context.md under 400 words, and one AGENTS.md per repo (CLAUDE.md just points to it).
- A post-turn hook in each agent, and one config file that maps light, deep and code-deep tiers to models.
- Known limits: required reading grew from about 6,000 to 20,000 words, and an append-only decisions file contradicted itself, so decisions are now add-or-replace.
- Who it’s for: developers who run more than one coding agent or bot daily and want shared memory without a vector database.
3. A proactive assistant as the main inbox: junk filtering, subscription cuts, podcast publishing · Workflow page
- Replaces: triaging hundreds of junk emails a day, auditing recurring costs by hand, and publishing each podcast episode manually.
- Results: stated results, from the author’s post:
- About $4,600 a month cut across 75 subscriptions and services.
- About 95% of episode publishing (Substack, YouTube, Libsyn) done automatically in 30 minutes instead of 2 hours.
- The same bot drafted and sent first pitches to about 70 prospects and 13 podcasts, and 5 hosts replied.
- What to connect: Grok Bot with email, calendar and the publishing accounts, plus routines for the follow-up. No cost is stated.
- The post ends with a plug for the author’s own add-on (Sidechannel), which is not needed for any of this.
- Who it’s for: solo operators running several small projects who want one assistant to handle the follow-through.
Quick hits
- Apple Intelligence reaches China with Qwen inside — the developer beta starts 4 November and the first batch ships with iOS 27.3 in December. The Verge · PoYo summary of Apple China Newsroom
- Alibaba’s Qwen powers Siri knowledge questions, Writing Tools and Visual Intelligence. It does not include the next-generation Siri AI announced at WWDC 2026.
- OpenAI plans a post-midterm policy blitz — Axios and TechCrunch report proposals for mandatory “speed limits” on the most powerful AI and community compacts for data-centre projects. TechCrunch · PoYo summary
- It includes a deliberate rebrand of data centres as “AI compute campuses”, and starts in the lame-duck Congress that opens 9 November.
- A quiet week for open weights, but a useful one — Nanbeige4.3-33B (Apache 2.0, 512K context, self-reported 80.9 on Terminal-Bench 2.1) and Liquid AI’s LFM3-1.7B (about 1 GB at 4-bit, for local agents) lead the list. Aiholics weekly roundup
- Also out: Tencent’s Hy-SimulMT-30B-A3B live speech translation and two OCR models. All benchmarks are vendor-reported.
- A four-day game side project, about 1,150 PRs — @poteto says every time he found an issue he asked his bot to write a lint rule against it. @poteto on X
- He also used TLA+ heavily to find bugs in the game’s state machines. It’s a concrete recipe for agent-heavy projects: turn each bug into a rule.
- Anthropic’s “don’t be cruel to Claude” policy gets a CNN explainer — the October usage-policy line bans “sustained and needless abusive or cruel behavior toward our models”. CNN via KSL
- The main enforcement is still Claude ending the conversation, with bans only in extreme cases. CNN notes that Anthropic is expected to IPO later this month.
Go deeper
- Nathan Lambert, “Was I wrong about open models & cyber risks?” (30 Oct) — a measured revisit of his “open-weight Mythos would’ve been fine” claim. Interconnects
- It uses the Epoch GLM 5.3 Firefox result, and takes the view that closed models are doing most of the damage so far.
- Talos on VOLTA — for the full prompt and failover design. Cisco Talos
- Frank Calegari, “Mathematics has changed. Let’s act like it” (guest post on Terence Tao’s blog, 1 Nov) — a mathematician’s argument for celebrating the 700-paper OpenAI release instead of mourning it. What’s new
- Parameter’s ChatGPT account-takeover writeup — for how a postMessage listener beat a sandbox that held. Linked from Simon Willison.