Proof, Provenance and the New Identity Problem

Claude’s Machine-Checked Mathematical Advance

An AI system has pushed a significant mathematical bound from 41.6 per cent to 67.2 per cent. It took 31 million output tokens, roughly 60 subagents and about 650 failed ideas to get there.

This is worth lingering on, because the achievement is more interesting — and more limited — than the phrase AI solves hard maths might suggest.

Anthropic reported on 10 August that an unreleased Claude research model generated a proof improving the known lower bound for the proportion of non-trivial Riemann zeta zeros lying on the critical line. That proportion was previously established at 41.6 per cent. The new argument raises it to 67.2 per cent.

That’s a substantial movement in a mathematical problem whose central conjecture, the Riemann hypothesis, has resisted proof for more than a century. But Anthropic is explicit about the boundary here: this work does not prove the Riemann hypothesis, and the company doesn’t expect the new method to provide a route to such a proof.

The validation process is as important as the number. Anthropic mathematicians reviewed the argument. A formal version written in Lean, a system that can mechanically check whether a proof follows its stated logical rules, passed machine verification. Outside mathematicians Brian Conrey and Dan Goldston also examined the result on short notice.

None of that is identical to conventional independent peer review, which hasn’t yet run its course. The argument builds heavily on existing mathematical literature, and its standing could change as more specialists scrutinise it. For now, the right description is a serious candidate research advance with unusually strong early checking, not settled mathematical history.

The scale of the experiment also tempers the easy automation story. Anthropic says it used about 31 million output tokens, around 60 subagents and hundreds of scripts. Roughly 650 earlier ideas failed. That sounds less like asking a chatbot one inspired question and more like operating a costly research programme that can search, test, discard and formally verify a huge number of possibilities.

For developers building scientific AI, the useful lesson is that the model didn’t need to become an infallible autonomous mathematician to produce value. It generated a novel candidate that survived formal checking and deserved attention from accomplished human experts. That’s already a meaningful capability.

My read is that machine-checkable fields may be where ambitious research agents earn trust first. When a system can expose its reasoning as a formal object, failed ideas can be discarded cheaply and the surviving result can face a verifier that doesn’t care how persuasive the prose sounds. Human judgement still decides whether the setup is relevant, whether hidden assumptions matter and whether the work contributes something worthwhile.

If the result holds up, the lasting advance may be less about this one bound than the workflow behind it: large-scale machine exploration funnelled into rigorous verification and specialist review. It’s powerful, expensive and still dependent on people who understand the domain.

Claude Adds Machine-Readable Marks

That research result shows why verification can’t stop at polished output. The same problem is now reaching everyday generated content.

Anthropic said on 11 August that Claude models launched in the European Union on or after 2 August 2026 will support machine-readable content marking from launch. Once supported, those marks are applied worldwide wherever the models are offered, rather than only to output produced inside Europe.

Anthropic says the coverage includes its API and products such as Claude, Claude Code, Cowork and Tag. Supported image files can also receive signed C2PA provenance metadata, giving compatible systems a way to inspect information about how a file was produced or processed.

The timing follows the European Commission’s guidance on Article 50 of the EU AI Act. The relevant transparency obligations have applied since 2 August. That regulatory pressure is beginning to show up as an engineering requirement inside model products, not merely as disclosure text tucked into terms of service.

But the meaning of a detected mark is narrow. Anthropic warns that it indicates content may have been processed by Claude. It doesn’t prove Claude authored the material, identify every other tool involved or establish a complete chain of provenance. Absence isn’t conclusive either. Paraphrasing, translation, extensive editing, very short output or stripped image metadata can weaken or remove the available signals.

Anthropic also hasn’t published complete detection documentation, so outside developers can’t yet fully assess how the signals behave across real production pipelines. Common operations such as copying text between applications or re-encoding an image could become important points of failure.

For teams distributing Claude-assisted material, provenance now needs operational ownership. Marks have to be preserved where possible, detection results interpreted conservatively and downstream users told what those results do and don’t establish.

My concern is that organisations could turn a useful compliance signal into a crude authenticity detector. A visible or machine-readable mark isn’t proof of deception, and an unmarked file isn’t proof of human origin. Used carefully, these systems can add evidence. Treated as a binary verdict, they could create precisely the false certainty that provenance tooling is supposed to reduce.

A Remote Worker Reaches Government

The limits of identity signals become much less abstract when the person on screen may not be the person doing the work.

An FBI cyber official has disclosed that investigators recently identified a North Korean remote IT worker employed by an unnamed United States federal agency. Todd Hemmen, deputy assistant director of the FBI’s Cyber Capabilities Branch, discussed the case at a government technology conference on 28 July. The disclosure was reported on 10 August.

The agency involved hasn’t been named. Nor do we know the employment mechanism, how long the worker had access, what systems were available or whether any information was stolen. The investigation is continuing, so there’s a large gap between the confirmed penetration and its possible consequences.

What the FBI has described is still striking. Hemmen said artificial intelligence is being used across these schemes: to create resumes and identity documents, get through video interviews, participate in meetings and conduct research. In other words, AI isn’t only improving a single fake artefact. It can support a consistent-looking identity across several stages of recruitment and employment.

That weakens a familiar security assumption. A convincing video call, plausible paperwork and competent work samples may all come from the same coordinated operation without establishing that the claimed worker is genuine. Passing an interview is only a snapshot, while the security exposure continues for as long as the person has credentials, devices or access to internal systems.

For government bodies and contractors, remote-worker verification therefore belongs alongside continuing personnel and supply-chain security controls. The strongest approach, in my view, is correlation over time: whether identity evidence, device information, location patterns and working behaviour remain consistent. No single signal carries enough weight on its own, particularly when generative tools can reinforce the others.

That doesn’t justify indiscriminate surveillance of remote staff. It does mean organisations handling sensitive systems need proportionate checks, clear escalation paths and verification that extends beyond recruitment. The practical shift is from asking whether someone passed an identity check to asking whether the evidence continues to support the same identity throughout the relationship.

A Logistics Breach Spreads Through Customers

Now consider a different kind of inherited risk: your systems stay online, but a company moving your goods is breached.

CEVA Logistics confirmed a cyber intrusion affecting part of its European contract-logistics operation. The company said the operational impact was limited to eight warehouses and that its other global systems were unaffected. Even with that boundary, disruption and breach notifications have spread across customers in retail, finance and gaming.

Reportedly exposed information includes customer names, postal addresses, email addresses, phone numbers and shipping details. The Dutch data-protection authority had received breach reports from ten organisations connected to the incident. The complete number of affected people remains unknown.

CEVA also hasn’t disclosed the initial access method, whether an extortion payment was demanded or the final recovery timeline. Those unknowns make it difficult to assess whether the eight-warehouse limit represents the finished scope or simply what has been confirmed publicly so far.

The episode illustrates why logistics providers occupy two risk categories at once. They process customer data, sometimes across several brands, and they also sit directly inside fulfilment operations. A compromise can therefore expose personal information while delaying the movement of physical goods. A customer organisation doesn’t need to suffer a direct network intrusion to inherit both problems.

For businesses relying on shared fulfilment, my takeaway is to map logistics firms as operational dependencies as well as data processors. Contracts and contingency plans need to address which customer fields are genuinely necessary, how quickly an incident is reported, what evidence is supplied and whether an alternative fulfilment path exists.

Data minimisation is especially practical here. Information that never reaches a logistics platform can’t leak from it. Where collection is unavoidable, organisations need to understand retention and onward sharing rather than treating shipping data as harmless administrative detail. The CEVA incident shows how eight affected facilities can create consequences across a much wider customer network.

Social-Media Design Claims Move Forward

There’s also a legal fight shifting attention from what people post to how platforms keep them engaged.

The Ninth US Circuit Court of Appeals has denied an attempt by Meta, TikTok, Snap and Google to pursue an early Section 230 appeal in consolidated litigation alleging that their product designs contribute to youth addiction and harm.

The proceedings bring together claims from individuals, governments and school districts. The plaintiffs are focusing on features such as recommendation systems, engagement-oriented design and warnings, rather than relying only on arguments about harmful material posted by users.

Section 230 is the legal protection platforms frequently invoke when claims concern user-generated content. The companies wanted an immediate appellate review of how that defence applies here. The court treated such an appeal as premature, allowing the underlying litigation to continue.

That procedural distinction is crucial. The ruling doesn’t establish that the platforms caused addiction, doesn’t decide that their designs are legally defective and doesn’t settle the ultimate reach of Section 230. Liability, possible damages and the defence itself remain unresolved.

What changes now is the pressure of continued litigation. The companies remain exposed to discovery and the prospect of trial, where internal evidence about product choices, recommendation systems and assessments of harm could become central. That can matter to corporate behaviour well before a final verdict, although the eventual outcome is still open.

For the public, the significant development is that courts are at least allowing a design-based theory of accountability to proceed further. My read is that the legal boundary around platforms may increasingly depend on whether a claimed harm comes from third-party speech or from the platform’s own optimisation and product decisions.

That’s not yet a winning theory; it’s a theory that survived this attempt to interrupt the case. The next stages will test whether plaintiffs can connect particular design choices to legally recognised harm, and whether the platforms’ statutory protections still apply when the disputed conduct is framed as their own product architecture.

Spotify Separates Synthetic Personas from AI Music

One platform is drawing a more precise line around AI: not how a song was made, but who appears to be presenting it.

Spotify announced on 11 August that profiles it identifies as AI personas will receive a visible label. Those profiles will also be excluded by default from editorial, algorithmic and personalised recommendations unless a listener follows them.

Self-disclosure through Spotify for Artists began on 11 August, with the public profile labels due to appear from mid-September. Spotify also plans to review profiles that meet audience thresholds it hasn’t disclosed. Creators will have an appeal process, although the company hasn’t yet explained enough to judge how effective that process will be.

The classification concerns the persona’s claimed identity. It doesn’t declare that the music itself was generated by AI. That distinction leaves room for artists to use AI-assisted production without automatically being treated as fabricated public figures, while giving listeners information when the performer identity is synthetic.

For listeners, the immediate benefit is a clearer identity signal. For operators of synthetic personas, the larger consequence may be reduced discovery: recommendation systems are a major route to an audience, and exclusion by default changes the economics even if the music remains available.

I think Spotify has chosen a more defensible boundary than trying to classify every use of AI in music production. Creative processes are mixed and difficult to audit, whereas presenting a fictional performer as a real person raises a more specific transparency issue.

The risk lies in enforcement. Spotify hasn’t published its review thresholds, detection accuracy or expected false-positive rate. A policy that affects recommendation access needs reliable classification and a workable appeal path. We won’t know whether it has those qualities until labels begin appearing in September and contested cases reveal how the system behaves.

Aptoide Enters Google Play

A small but concrete crack has appeared in Google’s control over Android app distribution in the United States.

Aptoide says Google has approved its Aptoide Games storefront for distribution through Google Play under the new Play Catalog Access program. Independent reporting describes it as the first major competing app store to return to Google Play through that route.

The scope is narrow. The storefront is focused on games, it’s available only in the United States, and it still operates under Google’s program rules and review. Some of the claims about the launch and its competitive significance also come from Aptoide’s own press release, so adoption and commercial impact need to be judged separately from the company’s framing.

Even so, the route is notable. A user can obtain a rival storefront from inside Google Play, rather than first changing device settings and sourcing an installer elsewhere. The change follows court-driven pressure on Google’s control of Android distribution and turns that pressure into a visible alternative on the platform itself.

For Android game developers, another storefront could eventually create another route to users and another set of commercial terms. But developer participation, user adoption, those terms and any expansion beyond US games haven’t been established.

My assessment is that this is evidence of platform opening, not proof of a competitive marketplace. A rival store that remains admitted through the incumbent’s program is different from fully independent distribution. Still, practical competition often begins with constrained access rather than a dramatic break. The test will be whether Aptoide can attract enough developers and users to exert pressure on discovery, fees or policies — and whether Google permits the model to expand.

What Changes for You

For Linux users, one persistent platform gap has finally narrowed, although the new option still carries a preview label.

OpenAI launched a worldwide preview of its dedicated ChatGPT desktop application for Linux on 11 August. It provides access to ChatGPT, ChatGPT Work and Codex, giving Linux-based users an official desktop surface instead of leaving the browser and unofficial packaging as their main options.

The supported list is specific: Ubuntu 24.04 and 26.04 LTS, Debian 13, and Fedora 43 and 44. Other Linux distributions derived from or adjacent to those systems may work, but OpenAI isn’t guaranteeing compatibility.

For individual developers, the immediate change is straightforward. ChatGPT and Codex can now sit in an official Linux desktop application alongside the rest of a development environment. That removes an awkward platform disadvantage for people whose primary workstation doesn’t run Windows or macOS.

The limitation is maturity. OpenAI is calling this a preview, not a general-availability release. Stability, behaviour under enterprise management and compatibility beyond the named distributions remain uncertain. Organisations managing Linux fleets will need to test those characteristics before treating the application as a dependable workstation component.

My practical read is that this is ready to evaluate, but too early to make foundational. Linux builders now have official access and a clearer support boundary; enterprise teams still need evidence on updates, deployment and reliability before standardising around it.