AI News Today — November 3, 2026: Anthropic Warns Self-Improving AI Is Near, Muse AI Pays $1.2 Million in Bug Bounties, and ChatGPT Merges Chat and Work

The big picture

Today's AI news is about speed — and who controls it. Anthropic told investors in an amended IPO filing that AI-assisted research may soon accelerate faster than its own ability to evaluate what it builds, while Meta put a price tag on agent security: about $1.2 million in bug bounties paid to keep Muse safe, including $300,000 for one flaw found before launch. The product news points the same direction — Anthropic made long Claude sessions roughly a third faster to start, OpenAI is collapsing ChatGPT's separate work surface into one tab, and Apple's watch can now rewind and recap your conversations. No new flagship model launched today. The systems around the existing ones are getting faster, more merged, and harder to supervise.

What happened today

Anthropic warns self-improving AI may be imminent — in its own IPO filing

What happened: In an amended S-1/A filed with the SEC on November 2, Anthropic added new language to its Risk Factors section, TechCrunch reports. The filing says the company's increasing use of AI to accelerate its own research "may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures," and admits it has in the past failed to notice a significant acceleration until after it happened. Its key sentence: internal estimates for future models — not current ones — "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI,'" the threshold tied to recursive self-improvement. Anthropic is explicit about the limits: it has not designated any current model as crossing that threshold, it is still assessing the data, and it says it does not anticipate being able to safely let Claude fully autonomously design and build successive frontier models without meaningful human supervision — adding that it is considering policy updates to keep humans in control of frontier development. One detail from wider coverage of the filing puts the acceleration in concrete terms: in September 2026, Claude reportedly authored over 90% of contributions to Anthropic's own codebase.

Who it matters to: Anyone who builds on, invests in, or regulates frontier AI. This is the first time a major lab has put a near-term recursive-self-improvement warning in a securities filing — a document where statements carry legal weight — rather than in a blog post. It is also, as TechCrunch notes, a filing that doubles as marketing to IPO investors, so the claim cuts both ways and should be read with that tension in mind.

What you can do: Nothing changes about the Claude you use today — no current model has crossed the threshold, by Anthropic's own statement. The practical move is governance, not panic: if your company is letting agents write and ship code with shrinking human review, Anthropic's filing is a well-timed prompt to check where your own review gates actually are before the tooling gets faster again.

Learn more: TechCrunch on Anthropic's amended S-1/A and the RSI warning

Muse AI: Meta has paid $1.2 million in bug bounties, including $300,000 for one pre-launch flaw

What happened: Meta's Muse team disclosed the scale of its security bounty program on November 2. David Singleton said Muse has now paid out more than $1 million — about $1.2 million in total — across more than 630 submissions, of which 64 were valid findings, RuntimeWire reports. The largest single payout was $300,000 for an authentication flaw found four months before Muse launched, which had exposed internal testers' tokens and conversations. The program also prices agent-specific attacks directly: a prompt injection carried out against a single user carries a ceiling of about $130,000. The disclosure lands a day after CNBC's report — covered in yesterday's briefing — that JPMorgan Chase, Charles Schwab, Amazon, Kohl's and Neiman Marcus have moved to block personal agents like Muse from their platforms.

Who it matters to: Anyone deciding how much access to hand a personal agent. Muse asks for the keys to email, accounts and purchases, so its security failures become account failures — and Meta is now showing its working: hundreds of researchers probed the product before and after launch, and the worst bug found would have exposed exactly the tokens and conversations users are asked to trust it with. The banks blocking Muse and Meta paying seven figures in bounties are two sides of the same negotiation over agent access.

What you can do: Treat the bounty numbers as a maturity signal, not a safety guarantee — 64 valid findings means the probing worked, not that the surface is clean. Before connecting an agent to email or payment accounts, check what it can do without asking you first, and keep approval switched on for sends and purchases. If you run a site or service, note Meta's pricing lesson: prompt injection against one user is now a six-figure bug class, so test your own agent integrations for it explicitly.

Learn more: RuntimeWire on Muse's $1.2M in bug bounties and the $300K pre-launch flaw

Anthropic opens message threads in public beta, cutting Claude Code waits on long sessions by about 35%

What happened: Anthropic opened message threads on the Messages API in public beta on November 2. Until now, an agent using the API resent the entire conversation — including tool results — on every turn. With threads, the conversation is cached on Anthropic's servers between turns and the client sends only the new messages, according to coverage of Anthropic's launch post. The feature is already live server-side in Claude Code for most users, with no setup required: on requests between 500,000 and 1 million tokens, Anthropic says the median time to first token drops by about 35%, and launch figures reported alongside the release put the median request shrinking from roughly 800 KB to 1.8 KB. For developers building their own agents, threads require a beta header and are billed per token at standard rates — the full conversation is still billed, only the resending stops. One design difference from rivals matters for privacy: OpenAI's and Google's equivalent chaining features keep state for weeks, while Anthropic keeps thread state in memory, expiring after about five minutes and deleted within an hour.

Who it matters to: Developers running long agent sessions — coding agents above all, where a single session can mean hundreds of API calls against a near-million-token context. A 35% cut in the wait before the model starts answering, multiplied across every turn, is a real productivity change, not a benchmark cosmetic.

What you can do: If you use Claude Code, you already have it — long sessions should simply feel less sluggish from today. If you build on the Messages API, enable the threads beta for long-running agents and measure your own time-to-first-token before and after; if your agent holds sensitive conversations, the five-minute, in-memory retention is a genuine point in Anthropic's favor compared with month-long state storage elsewhere.

Learn more: Coverage of the message threads beta and its Claude Code numbers

OpenAI is merging ChatGPT's Chat and Work tabs into one experience

What happened: OpenAI emailed users on November 2 that it will merge ChatGPT's Chat and Work tabs into a single experience, TechCrunch reports. The merged surface is meant to let a user "move from exploring an idea to having ChatGPT carry it through, all in one place" — bringing in files and plugins and asking ChatGPT to build a website, create a presentation, or act on the web while the user reviews and guides the work. Codex stays a separate experience in the desktop app. The rollout starts with a select group of users for feedback before going wide; OpenAI gave no dates. The email also confirms the usage change previewed in yesterday's briefing: with this transition, ChatGPT and Codex will draw tokens from the same allowance pool. OpenAI says the change won't impact most users, but that some may need to wait for resets more frequently depending on usage and model. TechCrunch notes rival Anthropic merged Claude's chat and Cowork functions into one screen in September — both labs are converging on one surface that routes work automatically instead of asking users to pick a tab.

Who it matters to: Anyone who works in ChatGPT daily, especially on paid plans. The tab you work in is about to disappear as a choice — which is simpler — but the shared token pool means heavy Codex sessions and heavy ChatGPT sessions now compete for one budget, as detailed yesterday.

What you can do: If you're in the early group, expect the interface to change without warning and don't build muscle memory around the Work tab. Either way, keep the baseline habit from yesterday's briefing: check Settings → Usage now, while chat and coding are still metered separately, so you can tell whether the merged pool changes your effective capacity.

Learn more: TechCrunch on the ChatGPT Chat and Work merge

Apple Watch learns to rewind and recap conversations, as Audio Intelligence enters beta

What happened: Apple's Audio Intelligence features for the Apple Watch Series 12 began beta testing for developers in the iOS 27.3 update, TechCrunch reports. The opt-in set has four parts: sound recognition for deaf and hard-of-hearing users (sirens, alarms, doorbells), Shazam-based song identification, Live Rewind — double-press the Digital Crown to see the last 15 seconds of a conversation as text — and Siri Recaps, which summarize a conversation into a short sentence and bullet points covering things like dates, deadlines and prices. Apple's privacy design is the story inside the story: a chime and animation signal when a rewind is captured, snippets delete automatically after 30 seconds if not saved, recaps delete within seven days if not kept, no raw audio is saved, speakers are not identified, and recaps are designed to omit sensitive information such as government IDs and financial or authentication data. Processing runs through the paired iPhone and Private Cloud Compute, with recap data deleted from Apple's servers immediately after creation, per Apple.

Who it matters to: Anyone who has watched the backlash to always-listening AI wearables — and anyone who takes notes for a living. Apple is making the opposite bet from the AI-pin era: recording features that announce themselves, expire by default, and keep almost nothing. Whether users trust a conversation recap on a mainstream watch will say a lot about where the line on ambient AI actually sits.

What you can do: Nothing yet unless you're a developer on the iOS 27.3 beta — and if you are, test the consent signals, not just the summaries: the chime, the animation and the 30-second expiry are the product. Everyone else can note the pattern for any AI recording tool: opt-in, visible capture, short retention, no raw audio kept. Tools that can't match that checklist deserve harder questions.

Learn more: TechCrunch on Apple Watch Audio Intelligence beta testing

Gemini Spark: no new Spark announcement — the Gems freeze is now 14 days out

What happened: Our daily priority check found no new Gemini Spark announcement. Spark remains folded into the Gemini app as Task mode — the Chat | Task switch, event-triggered multistep tasks across connected apps and the web, up to 15 tasks at once, for Google AI Pro and Ultra subscribers with a personal account, Keep Activity on and Personal Intelligence opted in, everywhere Gemini Apps are supported except the EEA, the UK, Switzerland and Nigeria, per Everpod's guide, checked against Google's help pages as of November 1. What is moving is the calendar underneath it, and today it is close enough to act on: personal-account Gems become read-only on November 17 — 14 days from today — and are permanently deleted on December 17. Google will auto-convert Gems to Skills, and the Skills format is the portable SKILL.md file Gemini now shares with Claude, but auto-conversion drops attachments over the 100 MB per-Skill cap, largest first, and some formats — legacy Office documents and archives among them — do not carry over at all. Workspace accounts migrate later, in March 2027 for business and nonprofit accounts and June 2027 for education.

Who it matters to: Anyone with saved Gems on a personal Google account — in two weeks the setups feeding your Spark tasks stop being editable, and in six weeks the originals are gone. Oversized or oddly formatted Gems are exactly the ones auto-conversion can quietly damage.

What you can do: Do this week what yesterday's briefing advised, because the deadline is now a fortnight away: open each Gem you depend on, copy its instructions into a plain SKILL.md file you keep yourself, and download any attachments over the 100 MB cap or in unsupported formats. That single file then works in Gemini's new Skills — and, because it is the shared format, in Claude as well.

Learn more: Everpod's Gemini Spark (now Task) guide, based on Google's help pages

Also worth knowing

• Anthropic's Claude text watermark expands to four older models today, November 3 — the change previewed yesterday takes effect on the Claude Platform for Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, with the caveat that on the 4.6 models the mark applies only at default generation settings. The Amazon Bedrock rollout may lag a few days.
• Goldman Sachs has pulled advanced AI coding tools back from most traders and bankers, according to reports summarized in The Neuron's November 3 roundup: after months of experimentation, tools like Claude Code are now largely reserved for engineers, with the bank citing productivity and risk-management concerns — a notable reversal from earlier praise for staff-built prototypes.
• WIRED uncovered more than 400 deceptive apps and websites linked to a network that disguised tools for generating sexually abusive deepfakes as ordinary services, including pet-portrait generators, per the same Neuron roundup: the apps had over 350,000 collective downloads reported, Apple removed 19 apps after WIRED contacted it, and Meta reported extensive enforcement against associated advertising accounts and pages.
• OpenAI and Anthropic stood up dedicated election-misinformation teams for today's U.S. midterms, per Semafor via The Neuron — response operations to detect misuse and handle misleading voting information as voters increasingly ask chatbots election questions. It is the first real operational test of both labs' election playbooks this cycle.
• Elon Musk throttled Grok Bot's limits, citing cost: users who found their weekly limits suddenly maxed out — some jumping from 50% to 100% with no overnight use — were not hit by a glitch. Musk confirmed on X that "the cost of serving Grok Bot was unsustainably high relative to revenue," and has not yet said how limits change going forward, TechCrunch reports.
• The New York Times and other publishers sued Brave Software, alleging the browser maker scraped their journalism without permission, circumvented paywalls and supplied access to AI companies, per The Neuron's roundup. The allegations have not been adjudicated; the suit adds another front to the publishing industry's licensing fight with AI search and data suppliers.
• Y Combinator's Garry Tan spends up to $200,000 a year on AI tokens — $100,000 to $200,000 by his own account, in a conversation with Walter Isaacson covered by TechCrunch and The Neuron. A useful ceiling marker for what unrestricted agent use actually costs a power user.
• Google took Dreambeans to India, its first market outside the U.S., with support for 13 local languages, TechCrunch reports. The experimental app turns activity from connected Google services into personalized illustrated stories — and its U.S. downloads fell 84% month over month in October, per Sensor Tower estimates shared with TechCrunch, a cautionary data point for proactive, prompt-free AI products.
• AGIBOT unveiled its GO-3 Preview embodied foundation model in Shanghai on November 3, claiming first-place rankings on the RoboDojo, RoboTwin and RoboColiseum robot benchmarks and promising a full open-source release by the end of 2026, per its PR Newswire announcement. Vendor-reported benchmark claims, as always, await independent replication.
• OpenAI also introduced LifeVal, a benchmark grading AI on everyday errands — finding childcare, budgeting, navigating government paperwork — built from time-use studies and a survey of 3,995 people, and announced its Grid Protectors program, under which a Georgia facility in development could make up to 1,000 megawatts available to the grid during emergencies, per The Neuron's roundup.
• Squarespace is laying off 19% of its employees, per TechCrunch's feed — another large restructuring at a web platform as AI builders and agents absorb work that once needed bigger product and support teams.

One practical takeaway

Price the access you hand an agent — before someone else prices it for you. Today's three biggest stories are the same lesson at three scales: Meta has paid $1.2 million to people who found ways into Muse, five major banks and retailers have decided Muse does not get in at all, and Musk just cut Grok Bot's limits because serving an agent cost more than it earned. Agents are no longer free, harmless or automatically welcome — every connection you grant one, and every allowance you build a workflow on, now carries a real cost and a real risk decision. This week, audit yours: list what your agents can touch, turn approvals back on for anything that sends or spends, and save local copies of the setups — Gems included, 14 days before they freeze — that you cannot afford to lose.

Sources

TechCrunch (Anthropic's amended S-1/A and the RSI threshold warning, November 2; the ChatGPT Chat and Work tab merge email, November 2; Apple Watch Audio Intelligence beta; Grok Bot limit throttling; Garry Tan's token spending; Dreambeans in India; Squarespace layoffs); RuntimeWire (Muse bug bounties — David Singleton's disclosure via muse.ai and X, November 2); Anthropic's message threads launch post via Pasqualepillitteri and JLS's blog (public beta, 35% time-to-first-token improvement, November 2); The Neuron's November 3 roundup (WIRED's 400+ deceptive deepfake apps; Goldman Sachs coding-tool restrictions via The Decoder; OpenAI and Anthropic election teams and OpenAI's Abu Dhabi agreements via Semafor; the New York Times publishers' suit against Brave via The Verge; OpenAI LifeVal and Grid Protectors); Everpod's Gemini Spark guide, checked against Google's Gemini Apps Help pages as of November 1 (Task mode requirements and availability); AGIBOT's GO-3 Preview announcement via PR Newswire/AAP, November 3. Summaries are original; follow the links for full reporting.