| Zymbos Intelligence · Friday 23 October 2026 | ||
| ||
|
A note from Milwaukee This week's issue comes all the way from Milwaukee, USA, which is why it reaches you a little later than usual. Tonight the Milwaukee Brewers host the Tampa Bay Rays in Game 1 of the World Series, the first time they have reached it since 1982, as Wisconsin Public Radio reports. There is a lot of excitement among my work colleagues here. | ||
When a convincing fake costs almost nothing to make, trust moves from what you say to what someone else can check. This week showed that shift from five directions. Meta's draft business terms put the risk of AI mistakes on the customer. The UK's financial regulator wants AI failures and near-misses on record. The US government made public a ban that turns on where an AI model came from. Microsoft says every AI agent needs its own identity and a watched trail. And Anthropic has let outside evaluators look inside and publish what they find. The lesson for your own work is the same in every case: make it cheap to verify. | ||
| ||
Contracts · Global Meta's draft terms: if its AI gets it wrong, that is on you Meta's proposed commercial terms for business customers, due to take effect on 30 October, make the customer responsible for the risk in what Meta's AI produces, according to CIO.com. The draft also makes customers answer for their partners and contractors, and gives Meta broad rights to crawl and use business content. Meta told CIO.com the crawling applies only to its advertising and business products, and that site owners can use a robots.txt file, the standard notice telling automated crawlers what they may read, to signal their preferences. The Register reports that the terms do not commit Meta to honouring that notice. Meta had already dropped explicit mention of AI training from the draft and said further changes were coming. The Register also reports that at least one million businesses use Meta's AI agents in WhatsApp and Messenger. Both reports cite Meta's published preview of the terms, which could still change before they start. McGann's TakeWhen the supplier will not stand behind its AI's output, checking that output becomes your job. Unverified work now has a price, and this contract says the customer pays it. Read the AI clauses in every business tool you accept, and put a human check on anything an AI agent sends in your name. | ||
Regulation · UK The FCA wants AI failures and near-misses on the record The Financial Conduct Authority (FCA), the UK's financial services regulator, expects firms to report significant AI failures and near-misses. In a speech on 19 October, Jessica Rusu, the FCA's chief data, information and intelligence officer, said existing rules already require firms to report major incidents, and that this must extend to AI. She also called for clear accountability, human oversight and escalation triggers. For AI agents, she described the assurance needed as continuous and conditional, because their behaviour can shift as they scale, under stress or when the underlying model changes. The FCA frames this as building on its current incident rules. The Register reports that the government has not yet committed to AI measures in the Cyber Security and Resilience Bill. McGann's TakeA near-miss log is the cheapest proof you will ever produce, because it is written while the facts are fresh. Rebuilding the story after a complaint costs far more and convinces far less. If AI touches regulated work, add AI failures and near-misses to your existing incident register, and expect supervisors to ask for it before any formal rule arrives. | ||
Procurement · US Where an AI model comes from now decides whether Washington can buy it US federal civilian agencies must stop buying technology from DeepSeek, the Chinese AI developer, and remove it from their systems, along with third-party products built on it, under a Department of Homeland Security (DHS) order made public on 19 and 20 October, according to TechCrunch and the Associated Press. The order itself is dated 24 July; only its publication is new. It excludes Defense Department agencies, and agencies can ask for exceptions. DHS cited the risk of sensitive government data being transferred to China. China objected and said it would take 'all necessary measures' in response. The reach is the point. Products built on DeepSeek's models are covered as well as DeepSeek's own app, so a supplier's brand tells a buyer little about the model underneath. Any agency buying software now has to know what sits inside it. McGann's TakeWhere a model comes from is now a question buyers ask before they sign. If you sell to government or build on open models, keep a list of the AI models inside your products, the way you already list software components. A vague answer to that question will start costing sales. |
Security · Global Microsoft says AI agents need watching like any insider Microsoft has told business customers to treat AI agents, software that carries out tasks on its own, as potential insider threats: monitored all the time and never trusted by default, TechCrunch reports. New Guardian Agents, part of a programme called Project Perception, supervise other agents. Microsoft Defender, its security product, will detect up to 60 types of agent, including Claude Code and Codex. Each agent gets permissions for the task in hand, and risky actions need human approval. In its own post, Microsoft describes giving each agent an identity, policies and permissions, so security teams can trace what happened and why. A Microsoft security executive said frontier models, the most capable AI systems, cannot be trusted whoever makes them. This is Microsoft's own account of its products, and nobody outside the company has yet tested whether they work. McGann's TakeAn agent that can send, buy or change things needs the same paper trail as an employee who can: an identity, limited permissions and a record of what it did. That record is provenance for actions. If you cannot say which agent did what, 'the AI did it' becomes the least useful sentence in the room. | ||
Assurance · US Anthropic lets outside evaluators inside, with the right to publish Anthropic, the company behind Claude, is embedding six evaluators from METR (Model Evaluation and Threat Research), an independent group that tests AI systems for risk, with access roughly on a par with its own risk teams, according to Anthropic. They get company hardware, office space and access to internal systems; Ars Technica, carrying Financial Times reporting, calls it employee-like access. METR can publish what it finds and Anthropic cannot veto publication, although it can require redactions in categories agreed in advance, and it decides what falls inside them. A second group, SecureBio, is reviewing the chemical and biological weapons chapter of Anthropic's August risk report. Anthropic describes the arrangement as voluntary, and ZDNET notes that third-party testing is not yet required by law. The terms are Anthropic's own account. METR's first published findings will show how the arrangement works in practice. McGann's TakeA claim another party can inspect is worth more than a claim you can only repeat. The telling terms here are the limits: what the evaluator can see, what it can publish and who decides the exceptions. Ask your own AI suppliers those three questions before you accept 'independently tested' as an answer. | ||
| ||
This Week's Analysis The Deepfake Discount A polished document used to earn some trust on sight. A clean report, a confident voice note, a sharp headshot: the effort behind them signalled that someone real stood behind the work. That signal is going. When a convincing fake costs almost nothing to make, polish tells a reader very little about where something came from. So they check, or they doubt. Call that the deepfake discount: the trust your work loses when a stranger cannot cheaply confirm where it came from and who stands behind it. My claim: the people and firms who do well from here will be the ones whose work is cheap to verify. Signed where the format allows, sourced where a claim depends on it, and consistent enough that a stranger can follow the trail. Three reasons. First, the discount is already being charged. Hiring teams check references, editors ask for the original file and buyers ask where a product came from. AI has raised the volume of convincing material, so those checks now land on more of what you send. This week the bill showed up in writing. Meta's draft terms (story 1) leave the customer carrying the risk of unchecked AI output. US federal agencies (story 3) must find and remove any product built on DeepSeek's models, which means knowing what sits inside the software they buy. In both cases, whoever cannot show what is inside the work pays. Second, proof is cheap for the sender and expensive for the receiver. Keeping your sources, your approved version and your sign-off takes minutes while the work is in front of you. Rebuilding that trail later, as a client, an auditor or a regulator, takes hours, and sometimes cannot be done. The FCA's call for a record of AI near-misses (story 2) is the same logic in regulatory form. When proof costs the sender so little and saves the receiver so much, the sender who supplies it wins the work. "Make your work cheaper to check than it is to doubt." Third, the standard exists; adoption is the gap. A Content Credential is a digital label sealed inside a file, like a signed delivery note. It records who signed the file, which app or device produced it and what was done to it. Any later edit is either recorded in the label or breaks the seal. The open standard behind it comes from the Coalition for Content Provenance and Authenticity (C2PA). This week's Tool on Trial reads one free in your browser. A credential shows where a file came from and whether it changed after signing. It does not prove the content is true. That limit is useful: it turns a vague claim into a narrow one a stranger can test. The strongest counter-argument is friction. Professionals are busy, the tools do not cover every format, and credentials can be stripped when a file is copied or shared. If proving takes longer than producing, most people will leave it to the platforms and hope the label holds. For low-stakes work, that is rational. It fails when the stakes rise: when a client disputes your advice, when a regulator asks for the log, when your contract says the AI's mistake is yours. So set a threshold. Sign and source anything that changes a decision, moves money, speaks for a person or will be reused without you in the room. For everything else, keep the source trail. That costs minutes per document. Platforms will add labels. They will not keep your sources, name your reviewer or date your approved version. That part stays with you. |
| |||||||||||||||||
Content Credentials Verify Provenance checker · Free · Runs in your browser
A free browser tool that reads the provenance label inside a file, and says plainly when it cannot vouch for it. What it is Verify is a free website where you drop in a photo, video, audio file or PDF and see any Content Credentials it carries: who signed it, which app or device produced it, what was done to it and which earlier files went into it. It is run by the Content Authenticity Initiative, an Adobe-led group, and built on the open C2PA standard explained in Deep Intelligence. There is nothing to install and no account to create. We tested it on a graphic carrying a credential from Anthropic's content-signing service. Verify named the issuer as Anthropic and the app as Anthropic Claude Content Signing, and listed the action recorded. It also marked the credential 'Unrecognized', because Anthropic's signing certificate is not on its list of known issuers, and warned that the file 'may not come from where it claims to'. That is the tool at its most honest, and at its most confusing. Who it suits Anyone who receives images, recordings or documents they might act on: hiring managers, editors, buyers, compliance teams. Also anyone who wants to see what signing their own work would show a stranger. Credentials are still rolling out, as the site itself warns, so many files you check will show nothing. Ratings Scores are from a desk review of the Verify page and its documentation, plus one hands-on test on 23 October 2026.
No account and no install. Drag in a file and the result appears in seconds. Sixteen formats, including PDF.
Credentials are still rolling out, so an empty result is the common case. Credentials can also be stripped when files are copied or shared.
Signer, app and edit history are shown in plain language. But 'Unrecognized' on a credential from a major AI company reads like a fraud alert, and the results page does not explain what a valid credential cannot prove.
An open standard backed by a cross-industry coalition, and at least one major AI company now signs files it produces, as our test showed. Recognition is patchy: Anthropic's certificate was not on the known list on test day.
Free, no sign-up, no paid tier. The cost of provenance sits in your workflow. Verdict The quickest way to see what provenance looks like in practice, limits included. Use it on anything you are about to act on that claims an origin. Read an empty result as 'not established', never as 'fake', and treat 'Unrecognized' as a prompt to check the source another way. Try Content Credentials Verify → Pricing verified on the live Verify page (verify.contentauthenticity.org, where the contentcredentials.org/verify address also leads), 23 Oct 2026. Free tier: £0 ($0), no account needed. Paid tier: none; no price, checkout or sign-up shown. Team tier: none. The real cost sits in adoption: keeping original files, using creation tools that sign their output, and agreeing who checks what. That is staff time, not a licence fee. | |||||||||||||||||
| |||||||||||||||||
The Verification Triage Scaled to the stakes · Works in Claude or ChatGPT Use this before you rely on a claim, image, recording or document that could change a decision: a supplier's figure, a screenshot in a dispute, a voice note asking for a payment, an AI-written summary of a contract. You tell the model what the item is, what you plan to do with it and what happens if it is wrong. It then separates three questions: who made the item, whether it has been altered, and whether the claim inside it is true. Many mistakes come from merging them. It orders the checks by stakes, cheapest first, and tells you when to stop. Good output is short: a table of checks with a status, the single next step and a plain verdict. It must not claim to have opened links or inspected files unless it actually did. For high-stakes calls, you stay the reviewer. It works in Claude or ChatGPT. You are a verification editor. Your job is to help me decide whether to rely on the item below, not to defend it. ITEM: [paste the claim, describe the image or recording, or attach the file] WHERE IT CAME FROM: [sender, website or platform, and date received] WHAT I PLAN TO DO WITH IT: [the decision or action it will feed] IF IT IS WRONG: [LOW / MEDIUM / HIGH, and why] Work through these steps in order: 1. Split the item into three questions and answer each separately: a) Origin: who made it, and can that be confirmed? b) Integrity: has it been changed since it was made? c) Truth: is the claim inside it accurate? 2. List the checks that would settle each question, cheapest first: the original file or source, the date, a second independent source, any Content Credentials or other provenance record, and confirmation through contact details I already trust. 3. Scale the checks to the stakes. LOW: one quick check. MEDIUM: origin plus one independent source. HIGH: the original, a primary record and independent confirmation before I act. 4. For any request to pay, share access or change bank details, tell me to confirm through a channel I already trust, never through details in the item itself. 5. Mark each check VERIFIED, CONTRADICTED or NOT YET CHECKED. Separate checks you actually performed from checks you are proposing. Rules: - Do not treat a missing credential as proof of fakery, or a valid credential as proof the claim is true. - Do not say you opened a link, inspected a file or contacted anyone unless you actually did. - Do not invent sources. Return: A. A table: check, status, what it shows. B. The single next check most likely to change the verdict. C. A verdict: RELY, RELY WITH CONDITIONS, or DO NOT RELY YET. D. Two lines I can keep on file: what I checked and when. Verify in proportion to consequence, sign everything you ship. |
| ||
Closing Perspective Signed by default Most of the provenance debate is about spotting fakes. The more useful question for a professional runs the other way: could a stranger confirm that your work is really yours? The first documents to carry credentials by default will be the ones most worth impersonating. Company press releases are the obvious candidate. Listed companies have the most to lose from a convincing fake announcement, journalists need to know a statement is official, and the communications teams who issue releases already control the tools they publish with. Once set up, signing every release costs them almost nothing. My prediction: by 31 December 2027, at least three FTSE 100 companies will attach Content Credentials by default to the PDF press releases they publish on their own websites. You will be able to check it yourself: download a release and drop it into this week's Tool on Trial. A one-off pilot does not count. Until then, start with one document of your own. Keep the sources, date the approved version and name the reviewer, then ask someone outside the work to trace where it came from. Reply with the last thing you shipped that a stranger could verify came from you, and the one piece of evidence they would check. John McGann Founder, Zymbos AI |
Zymbos Intelligence zymbos.ai You're receiving this because you subscribed at zymbos.ai © 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |