AI Street Journal for 9 September 2026
Google DeepMind says it has published predicted effects for every possible single-letter DNA change in the human genome. The resource is free for academic research, but it is explicitly not approved for clinical use.
The briefing
A genomics dataset arrives with grand scale and clear caveats, Anthropic users face a grubby account-security problem, and investors hand Cognition another very large cheque. In each case, the practical question is less magic and more who can use it, what it costs, and where the awkward limits sit.
AlphaGenome Atlas opens variant predictions to researchers
Google DeepMind says AlphaGenome Atlas offers precomputed predictions for about 9 billion single-letter DNA variants, with a new scoring system to rank likely impact.
Google DeepMind says AlphaGenome Atlas packages predictions for roughly 9 billion possible single-nucleotide variants, covering every single-letter DNA change in the human genome. The company says researchers can browse it through a website portal, use it through the AlphaGenome API, and access it as a skill in Google Antigravity. It also introduced an AlphaGenome Variant Impact, or AVI, score intended to rank variants by likely molecular effect.
The announcement matters because most genetic variants cannot be tested one by one in the lab. DeepMind says the atlas links each variant to thousands of predicted molecular effects, feature attributions explaining what drove the score, and a library of more than 2,500 recurring DNA motifs. It says the dataset is about 1 petabyte, making it larger than a casual download for the lunch break.
What the company says it can already do
DeepMind says external academic collaborators have used the resource in rare-disease work and population genetics. In one collaboration with researchers in the GREGoR Consortium, the AVI score was used to prioritise variants in unsolved rare disease research; the post says experimental screens then validated a DNM1-related splicing effect. In another example, Gareth Hawkes at the University of Exeter applied the atlas to UK Biobank whole-genome data and, according to DeepMind, found more non-coding associations by grouping rare variants by predicted molecular effects.
The company says the tool is available for non-commercial use through the website now, with commercial access on Google Cloud coming later. It also states clearly that AlphaGenome Atlas is not a substitute for medical advice and has not been validated or approved for clinical use.
For biology labs, the immediate change is access to a searchable ranking system for variants that would otherwise take years to sift through. The limit is equally important: these are model predictions for research use, not a diagnostic result a hospital can safely treat as settled fact.
Market signal
Anthropic users reportedly hit by stolen Claude sessions
TechCrunch reports that stolen Claude sessions were used to consume subscribers’ usage, while user complaints highlighted limited itemised visibility into where tokens had gone.
TechCrunch reported that at least some Anthropic customers had Claude usage consumed without their knowledge after account access was stolen. Its article centres on Grant De Swardt, an independent consultant in East Sussex, who told TechCrunch that his Claude Max account kept burning tokens while he was not using it. After he contacted Anthropic, the company suspended the account, invalidated sessions and server-side Claude Code tokens, and issued a partial refund, according to TechCrunch.
TechCrunch says Anthropic later told De Swardt that a compromised Claude session key had been used to mint unauthorised Claude Code OAuth tokens. The company reportedly said the evidence was consistent either with credentials or session data being taken without his knowledge, or with the account having been connected to an outside service.
What users could and could not see
The article says De Swardt asked for an itemised usage list but did not receive one. TechCrunch also cites Reddit posts, a GitHub report and emails users said they received from Anthropic. In those emails, Anthropic said it had become aware of a bad actor using common infostealer malware to steal Claude login sessions from people’s computers and then consume account usage.
According to TechCrunch, Anthropic signed affected users out, invalidated existing authorisations, issued some refunds and warned that their machines might have malware. The report also says Anthropic declined to comment when asked how users can identify misuse. That leaves an awkward gap between total usage and usable forensic detail for customers trying to work out what happened.
The practical consequence is simple enough: paid usage can reportedly disappear without a clear transaction trail, while customers may struggle to distinguish malware, stolen sessions and authorised integrations. For businesses that rely on Claude-based workflows, account hygiene and session management look less optional than they did before.
What to watch
Cognition raises 2 billion at 48 billion valuation
Cognition announced a $2 billion round at a $48 billion valuation, while TechCrunch says the financing underlines continuing investor appetite for AI coding companies.
Cognition announced that it had raised $2 billion at a $48 billion valuation, according to TechCrunch. The report says the round was led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir, and came four months after a previous fundraise at a $26 billion valuation.
TechCrunch frames the deal as evidence that venture investors still see room for multiple large AI coding businesses. Cognition told the publication that its annualised run-rate revenue had risen from $492 million to $900 million since its earlier round. The company did not explain how it calculated that figure, TechCrunch noted.
What the money says, and what it does not
The article compares Cognition with Cursor, which TechCrunch says had discussed fundraising at a $50 billion valuation before agreeing to sell to SpaceX for $60 billion. On that comparison, TechCrunch argues that Cognition is commanding a rich revenue multiple relative to Cursor’s earlier fundraising talks.
The same report also points to cost pressure. TechCrunch cites The Information as saying Cognition leases an Nvidia server cluster that costs hundreds of millions of dollars a year and that total cash burn could reach $800 million this year. It adds that Cognition, like Cursor before its sale, is training its own model based on open-source alternatives to reduce dependence on third-party models from OpenAI and Anthropic.
For customers, the announcement suggests AI coding tools will remain a crowded supplier market rather than immediately narrowing to one platform. For investors and rivals, it is a reminder that revenue growth still attracts capital, but infrastructure bills can be large enough to spoil the party if growth slows.
What to watch next
- For biology labs, the immediate change is access to a searchable ranking system for variants that would otherwise take years to sift through. The limit is equally important: these are model predictions for research use, not a diagnostic result a hospital can safely treat as settled fact.
- The practical consequence is simple enough: paid usage can reportedly disappear without a clear transaction trail, while customers may struggle to distinguish malware, stolen sessions and authorised integrations. For businesses that rely on Claude-based workflows, account hygiene and session management look less optional than they did before.
- For customers, the announcement suggests AI coding tools will remain a crowded supplier market rather than immediately narrowing to one platform. For investors and rivals, it is a reminder that revenue growth still attracts capital, but infrastructure bills can be large enough to spoil the party if growth slows.
The takeaway
For biology labs, the immediate change is access to a searchable ranking system for variants that would otherwise take years to sift through. The limit is equally important: these are model predictions for research use, not a diagnostic result a hospital can safely treat as settled fact.
The editor’s view
The practical consequence is simple enough: paid usage can reportedly disappear without a clear transaction trail, while customers may struggle to distinguish malware, stolen sessions and authorised integrations. For businesses that rely on Claude-based workflows, account hygiene and session management look less optional than they did before.