nikky

A Flaw in OpenAI’s Codex Put 97 Million Monthly Downloads at Risk

96.8 million. That is how many times the command-line version of Codex, OpenAI’s coding assistant, was downloaded in the 30 days to 10 November. Security researchers say one fake bug report could have let an attacker slip harmful code into a future update of it. OpenAI has fixed the problem and paid the researchers a first reward of $1,000.

The researchers work at the security firm Depi. On 11 November they published every step of the attack. The download count comes from the public statistics of the npm software library. Depi says users of Codex have nothing to patch.

How the attack worked

OpenAI keeps the code for Codex in a public project on GitHub, where anyone can file a bug report. A helper tool read each new report and suggested labels for it. That helper was the Codex assistant itself, and it received the text of the report mixed in with its own instructions. So a writer could hide commands in a report. This trick is called prompt injection.

Depi says it got the assistant to run hidden code, broke out of the closed space it works in and read secret access keys from the computer’s memory. With those keys it planted a harmful file in a store of files that automated jobs share. A more powerful job could then run that file. The last step was a trap that would give the attacker a key the npm library accepts for publishing new versions of Codex.

What OpenAI changed

Depi says OpenAI deployed a fix right away and paid a first reward of $1,000, with the report still under review. The highest rewards for the most serious bugs reach $100,000 before bonuses. The company raised its reward rates earlier this month.

Early on 12 November, Paris time, OpenAI replaced the assistant in that labelling tool with a short script. The script asks a model for labels with its tools switched off and only accepts answers in a fixed format. The change is visible in the project’s public list of changes.

There is one caveat. Depi ran the full attack on its own test copy, where the assistant was allowed to write files. OpenAI’s own helper had used a stricter read-only mode since May. The setting appears in the version of the file from that month. Depi does not say whether the attack would have worked in that mode, and it does not report any harmful update reaching users.

Its advice for everyone else is to wait a short time before tools install new updates automatically. Last month another team received a reward of $5,000 for a flaw in shared chat links.

Why it matters: AI assistants that read text from strangers can be talked into acting on it. Here the stranger only had to file a bug report, and the target was a tool that runs code on developers’ computers tens of millions of times a month.

The post A Flaw in OpenAI’s Codex Put 97 Million Monthly Downloads at Risk appeared first on DailyAI.

Japan Drafts AI Safety Rules That Would Carry Penalties

Japan’s main AI law has been in place since 2025, and it contains no penalties at all. Government advisers have now drafted a plan to change that for the most powerful AI systems. Companies that ignore an official order could face a penalty.

The draft was presented on Wednesday 11 November at a closed meeting of a working group under the Cabinet Office, the department that supports the prime minister. The government has published the documents from the meeting online. Japan passed its AI law in May 2025. Today the government can investigate a company and give it guidance, but it cannot force the company to act.

Four steps, from reports to orders

The new rules would not cover all AI. They would apply only where two things meet: AI above a certain level of performance, and a set of specific serious risks. Other AI would stay under existing laws and guidelines. According to the meeting documents, the draft has four parts.

  • Companies that meet the criteria must take measures to keep their AI safe from those risks.
  • They must report information about those risks to the government.
  • The government could formally recommend steps to a company that falls short.
  • If needed, the government could issue a binding order, with penalties for breaking it.

Members of the group named cyberattacks and attacks using chemical, biological, radiological, nuclear or explosive weapons as risks that need action. The rules would target companies that provide AI first, not people who use it. The group still has to decide which kinds of company are covered, from research labs to firms that sell AI services. Members also said the rules must work against foreign companies, so the burden does not fall only on Japanese ones.

Six companies took part

Six companies spoke to the group. Three of them were Japanese: the telecoms group NTT, the startup Sakana AI and SB Intuitions. The other three were Anthropic Japan, Google and OpenAI Japan. Each company was asked about AI risks, its own safety work and which rules it would support.

Penalties are expected to apply only to incidents involving advanced AI, according to the news agency Kyodo. The group aims to publish an outline of the changes, including any revision of the law, by the end of the year. The size of any penalty has not been set.

The meeting documents also compare the plan with rules in the US and the EU. New York and Illinois give the largest AI developers 72 hours to report a critical safety incident, and California gives 15 days. Under the EU AI Act, companies behind the most powerful models can be fined up to 3% of global turnover or €15 million.

New York opened registration under its own AI safety law this week. Australia also announced plans for an AI safety regulator.

Why it matters: Japan has so far relied on voluntary guidance for AI. If this draft becomes law, the biggest AI labs would face safety rules backed by penalties in Japan as well as in the EU and several US states.

The post Japan Drafts AI Safety Rules That Would Carry Penalties appeared first on DailyAI.

DailyAI Briefing: Anthropic’s IPO could raise $64bn, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Anthropic has put a price range on its stock market debut, and at the top it would raise more than $64 billion. New York gave the biggest AI labs until 1 January to register with the state, a small team used AI to prove key security properties of the code behind Signal in four weeks, ChatGPT opened its doors to sales chatbots run by businesses, and a new study shows freelance writing losing ground on Upwork.

Let’s dig in.

Anthropic’s IPO could raise $64bn

$64.35 billion. That is how much Anthropic would raise if it sells 825 million shares at $78, the top of a $70 to $78 range, before it pays the banks. The company set out the terms in a short announcement on 10 November. The shares will trade on the Nasdaq under the symbol ANT.

The updated prospectus has the most useful numbers. Revenue in the first nine months of 2026 was $33.24 billion, against $2.55 billion a year earlier. The operating loss shrank to $1.26 billion from $5.22 billion. The net loss of $104.0 billion looks alarming, but $94.3 billion of it is an accounting charge on convertible notes, not money the company spent. The full filing is on the website of the American securities regulator. CNBC reports that pricing is due on 18 November and trading on 19 November, with about 20% of the shares set aside for ordinary investors. We went through the numbers in detail. The banks will pitch sales that grew about 13 times in a year. The final price will show how much investors believe in that growth.

New York gives big AI labs until 1 January to register

72 hours. From 1 January, that is how long the largest AI developers will have to report a serious safety incident to New York’s financial regulator under the RAISE Act. The governor’s office opened the registration portal on 10 November. The law covers companies with more than $500 million in yearly revenue that train models using more than 10^26 computing operations.

Each covered lab must publish a safety plan, send the regulator a summary of catastrophic risks every three months and publish a transparency report with each major model release. Fines reach up to $1 million for a first violation and $3 million after that, according to reporting on the law. The state is also funding a testing centre at Binghamton University with more than 80 researchers. Our article lists every duty in the law. For now, the 72-hour reporting rule is the strongest part of the law.

AI wrote proofs for Signal’s code in four weeks

Four weeks and five people. That is what it took a Google DeepMind team, working with AI systems that wrote proofs in Lean, to check key security properties of libsignal, the library behind Signal’s encryption. The proofs run to about 6.5 million lines and cover the compiled machine code, not a tidy rewrite. The research post explains the method and its limits.

The fine print matters. The work covers one-to-one sessions only, so group chats, the server and the sealed sender feature are left for later, and the cryptographic building blocks are assumed rather than proved. Humans still chose what to prove and signed off the specification. As a test, the team put back two old bugs and showed that the security theorem then fails. We explained what was proved and what was not. The seL4 kernel took about twenty person-years to verify, so this is a big change in speed. Even so, the claim on social media goes further than the research paper does.

ChatGPT lets businesses add sales chatbots

1.2 billion weekly users. That is the audience OpenAI is offering businesses with a new beta that lets online shops in America add their own chatbot to ChatGPT. The launch post describes the new business profiles. Users can open the chatbot with one click from a ChatGPT answer, and brands can also pay to promote the same chatbot as an ad.

OpenAI has done some things well. Businesses do not see users’ chats, ads run on a separate system and appear only for free and Go users. The weak spot is the chatbot itself. It is built from instructions and a product feed that the seller supplies, and the only sign that it speaks for the seller is a label. We argued that these chatbots should say who they work for. A clear written statement at the start of every chat would cost OpenAI very little.

Freelance writing loses ground on Upwork

About 31%. That is how far the share of Upwork job posts asking for content writing fell between the first half of 2025 and the first half of 2026, according to a joint report by OpenAI and Upwork. Writing made up 14.5% of sampled ChatGPT conversations but only 1.3% of the job posts. The full report sets out the method and the caveats.

Video and animation went the other way, with 10.0% of job posts and just 1.0% of conversations. Jobs that involved AI also carried fixed budgets about 52% higher. The study compares percentages, not totals. It does not prove cause, and both companies have a stake in the result. Our report covers the other findings. Writers who rely on simple briefs should take note.

Signing off

Wednesday was about money and proof. Anthropic opened its accounts to the public, New York started a clock on AI incidents, and Google DeepMind showed how fast machine-checked mathematical proofs can now arrive. Meanwhile, chatbots are moving into shopping and into the work freelancers used to do.

See you tomorrow.

The post DailyAI Briefing: Anthropic’s IPO could raise $64bn, plus 4 more appeared first on DailyAI.

ChatGPT’s New Business Agents Should Admit They Are Salespeople

Companies can now add their own sales chatbot to ChatGPT. Users can open it with one click from an answer, and companies can pay to promote the same chatbot as an ad. The idea is sound. But a chat that speaks for a seller should say so in plain words, every time, and not rely on a small label.

This is a DailyAI analysis. The facts below come from the company’s own pages and from what its executives told the trade press. The opinions in it are ours.

The new tools for businesses

On 10 November, OpenAI opened Businesses in ChatGPT as a beta. Companies can now claim a profile, add product feeds with prices and stock levels, and see how often they come up in conversations. The pitch is reach. ChatGPT has 1.2 billion weekly users, the company says.

The bigger change is the Business Agent. Online shops and online service firms in America can build a chatbot that answers questions about what they sell. When a business appears in an answer, a user can open its profile and start a chat with its agent. That chat is labelled and kept apart from the main conversation. When the shopper is ready to pay, the agent sends them to the seller’s own website.

Brands can also pay to promote the chatbot as an ad. An eligible ad can carry a Start chat button that opens the chatbot, next to a link to the website. Benji Shomair, the company’s vice-president of monetisation, explained the link between the two in a press briefing. “Think of the business agent as the organic experience,” he said. “The sponsored agent is the paid experience, but the agent that sits behind it should be the same,” he added. His comments were reported by a trade magazine for the advertising industry.

What OpenAI got right

Much of the design is careful. Businesses do not see users’ conversations or their chats with a brand’s chatbot. They get grouped themes, and only once enough people have taken part. Shomair also said that claiming a profile does not move a business higher in the answers. In his words, the core product must stay objective, and a business should not be able to change how often it shows up in answers.

Ads are kept apart from the chatbot too. According to the engineers who built it, the ad system runs as its own service, with separate models and separate code from the one that writes answers. Ads appear only for people on the free plan and the low-cost Go plan, below the answer and marked as sponsored. They are already big business. Advertising in ChatGPT passed $1 billion in annualised revenue less than 200 days after launch, and more than 100,000 advertisers have used them.

Where the design falls short

The promise that ads do not sway answers applies only to the chatbot’s own replies. It does not apply to a Business Agent. Each one is built inside ChatGPT from instructions and a product feed that the seller supplies. Shomair said firms cannot bring a chatbot of their own. He described typical use as shoppers weighing up size, colour and price before looking at related purchases. Early tests of the paid version had shown “promising results”, he said, especially in the depth of the conversations.

That is a sales assistant. There is nothing wrong with a sales assistant. But people treat a shop clerk differently from a friend who knows the products, and here both speak in the same app, in the same chat style, a click apart. The company’s own example has a shopper asking a Business Agent about the differences between two options and which features fit their needs. A seller’s chatbot is not the right place to settle that question.

A label at the top of a chat window is easy to miss, especially on a phone. And the people most likely to miss it are the ones who already trust ChatGPT’s answers the most.

Three changes OpenAI should make

The fixes would cost little. First, every Business Agent should begin each chat by stating, in words, that it speaks for the seller and cannot give a neutral comparison with rivals. Second, a paid Business Agent should look different from a free one, so users can see when a brand has paid to reach them. Third, the rules these chatbots must follow on prices, discounts and refunds should sit in one public place that users can find, along with how they are checked.

Shopping through chatbots is growing fast. Earlier this month we looked at a claim about how many searches the chatbot now handles. We have also argued that large stores which block AI shopping agents will not hold out for long.

Why it matters: ChatGPT is becoming a place where people decide what to buy, and sellers can now talk to shoppers there. Users should always be able to tell, at a glance, when they have moved from an independent answer to a sales pitch.

The post ChatGPT’s New Business Agents Should Admit They Are Salespeople appeared first on DailyAI.

A Five-Person Team and AI Proved Signal’s Code in Four Weeks

Four weeks. That is how long a five-person Google DeepMind team, working with AI agents, took to produce machine-checked proofs of key security properties in the compiled code of Signal’s core chat library. The proofs run to about 6.5 million lines. Earlier landmark efforts on other systems took years of specialist labour, and usually needed code written to be proved.

The research post, dated 10 November 2026, sets out the method and its limits. The accompanying technical paper reports the numbers in more detail. The work is not a proof of the Signal app as a whole, and it is not a proof of other apps that use a related design. It is a proof of specified properties of libsignal, the reference library, for one-to-one sessions, taken down to the compiled machine code that actually runs.

What was proved

libsignal is the library that implements Signal’s encryption design. In plain terms, that design is meant to keep a message private even if a key is stolen later, and to recover after a device has been briefly compromised. Those two ideas are usually called forward secrecy and post-compromise security.

The team compiled the unmodified library — version 0.105.4, plus the related SPQR crate at version 1.6.0 — with the ordinary Rust compiler, into a small command-line tool running on 64-bit Arm Linux. The proofs are about that executable, not a cleaned-up rewrite. The verified library is about 560KB, or about 140,000 machine instructions.

They cover the PQXDH handshake, which is how two people start a private session, including a post-quantum step meant to stand up to a future quantum computer, and the Triple Ratchet, which issues a fresh key for each message. In the “symbolic” model, which treats the underlying cryptography as ideal, the paper also claims handshake secrecy against a later attacker with a quantum computer. Group chats, sealed sender and the server-side protocol are left as future work.

How the work was done

Formal verification means writing a precise mathematical statement of what the code must do, then proving that the code meets it. A single missed assumption can weaken the result, so people still have to choose what to prove and approve that statement.

The paper is blunt about the split. Humans selected the properties, resolved ambiguities and signed off the specification. Agents then wrote the proofs. Those proofs are accepted only if Lean, a proof-checking system, confirms them. The reliability of the result, the authors write, does not depend on the reliability of the proving agents, because Lean’s kernel checks the logic.

The research post says the proving agents were powered by Gemini Carbon. The proof work itself was done by five researchers over four weeks, each running agents on their own workstation. That is a short campaign by the standards of the field. The paper notes that the seL4 microkernel, a famous earlier verification, took twenty person-years, and that the CompCert C compiler took six. Both of those codebases were written with proof in mind. libsignal was not.

The proofs themselves can be long because they talk about machine code. About 6.5 million lines are proofs: 5.3 million about the binary, plus separate files for the two security models. A reader who wants to trust the headline claims still has to read the trusted parts: about 18,000 lines of specification in the repository, plus about 8,000 lines of the Arm machine-code model and the symbol names mapped onto the executable. The low-level cryptographic routines, two output functions and a few library calls are assumed by contract, not proved.

A check against old bugs

To test whether the statements were strong enough, the team reintroduced two bugs that researchers had previously found in libsignal, and that have since been fixed. In both cases they proved that the symbolic security theorem does not hold of the bugged binary. That is a useful check: a specification that still “succeeds” on known-broken code would be too weak.

One of those bugs stored a new session’s state, including a copy of a secret, after a prekey message had only decrypted under an older session. The other accepted a handshake whose base key had been mangled, creating a second session with the same shared secret.

What the result does not say

The authors are careful, and the limits matter. The proofs do not cover every part of Signal, every possible attack, or the cryptographic primitives themselves. Side channels are out of scope. The computational proof, which tries to bound a real attacker’s chance of success, gives no guarantee against a quantum adversary, and its bound becomes vacuous for multiple long sessions with many rounds.

Variants of the same protocol are used in WhatsApp, Meta’s Messenger and Google Messages, the paper notes. Those apps were not the artefact under proof. A proof about libsignal does not automatically transfer to every product that speaks a similar language.

Ehren Kret, Signal’s chief technology officer, is quoted in the research post as saying that rigorously verifying the protocol implementation “gives us even greater confidence in our code’s privacy guarantees”.

The proofs and specifications are published in an open repository.

Why it matters

Most people who send a private message never see the library that encrypts it. A method that can check the compiled code of that library, in weeks rather than person-years, is a practical change in how such software can be audited. It is not a certificate that Signal, or any other chat app, is free of flaws. It is a machine-checked argument about specified properties of one core library, produced quickly enough that other critical code might be next.

Most everyday chat products still have no such proof, and some popular tools do not encrypt messages end to end at all. An earlier report covered a consumer chat tool that sends texts without that protection.

The post A Five-Person Team and AI Proved Signal’s Code in Four Weeks appeared first on DailyAI.

Freelance Writing Work Shrank 31% as a Share of Upwork Job Posts

Content writing made up 14.5% of the ChatGPT conversations in a new study, but only 1.3% of job posts on the freelance site Upwork. The share of those job posts that ask for writing shrank by about 31% in a single year.

The figures come from a joint report by OpenAI and Upwork. The main findings were set out in a research summary published on 10 November. Upwork is an online marketplace where companies hire freelancers for short projects. The researchers compared 36,896 sampled ChatGPT conversations from January 2025 to June 2026 with the jobs that clients posted on Upwork. An AI classifier sorted the conversations, and OpenAI says no researcher read user messages.

Freelancing is not a niche. Upwork’s own survey this year found that 38% of skilled American knowledge workers do freelance work.

Where the work is moving

The full report finds that the share of job posts on the site asking for content writing fell by roughly 31% between the first half of 2025 and the first half of 2026. In plain terms, some clients may now draft blog posts, marketing copy and product descriptions with ChatGPT instead of paying a freelancer. The authors say firms “may be using AI directly for this task instead of hiring remote workers”, and they warn this could put pressure on freelancers who specialise in writing.

Video shows the opposite pattern. Video and animation made up 10.0% of Upwork job posts but just 1.0% of the ChatGPT conversations. Between early 2023 and late 2025, video job posts grew by about a third relative to the whole marketplace. Work on mobile apps follows the same pattern. It made up 7.8% of job posts and 0.2% of conversations.

The researchers checked 63 types of work. In 43 of them, tasks that were common in ChatGPT lost ground on Upwork, or tasks that were rare in ChatGPT gained ground. The other 20 went the other way, so ChatGPT use alone does not explain the shift.

Bigger budgets for AI projects

For similar work posted in the same month, fixed-price budgets were about 52% higher when the job involved AI. Freelancers on hourly contracts spent about 41% more time on such work in the first 90 days. Hourly rates were only 1.5% higher. That points to larger projects, not better pay per hour.

What the numbers do not show

A smaller share is not the same as fewer jobs. The report says its comparison does not measure overall growth on the site, and it does not prove that ChatGPT caused the change. Both companies also have a stake in how AI and freelance work are seen, and the study rests on their own data.

Earlier this month, another study found that people who use AI at work are putting in longer days. In a separate test, the best AI coding agent fully completed only 26.4% of real tasks.

Why it matters: Freelance writing is one of the first kinds of paid work where AI use shows up in hard numbers. Writers are losing ground on Upwork, while projects that involve AI are getting bigger.

The post Freelance Writing Work Shrank 31% as a Share of Upwork Job Posts appeared first on DailyAI.