The AI Brief

At an event in San Francisco on June 30, Anthropic's life sciences head Eric Kauderer-Abrams announced Claude Science, described as a public-beta workbench integrating more than 60 preconfigured scientific databases, tools, and connectors into a single research environment with provenance tracking on every result. The platform covers genomics, single-cell analysis, proteomics, structural biology, and cheminformatics, and can orchestrate cluster compute jobs while maintaining session memory across research workflows. Simultaneously, Kauderer-Abrams confirmed that Anthropic will run its own preclinical drug discovery programs, starting with neglected diseases that traditional pharmaceutical companies have largely abandoned.

Anthropic has not disclosed specific disease targets, team size, or budget for the drug programs, and has not clarified whether it intends to bring candidates to commercialization or partner out at the clinical stage. Among early Claude Science users named publicly: a UCSF researcher identified viral contamination in a dataset "within minutes, contamination that had gone undetected for a year"; the platform flagged 32 rare disease candidates for computational screening from 100 diseases analyzed in under an hour; and epidemiologist Stephen Francis at UCSF reported roughly a 90% speedup in glioma molecular epidemiology analysis. Novo Nordisk and the Allen Institute are among early adopters. Anthropic is offering up to $30,000 in API credits for approximately 50 research projects, with applications closing July 15 and recipients announced by July 31.

Novartis CEO Vas Narasimhan has said AI has the potential to cut drug development timelines from twelve years to seven or eight and double success rates from roughly 8% to 16%. Anthropic cites its public benefit corporation structure as the legal mechanism enabling it to pursue therapeutically important but commercially unattractive disease areas. No AI-designed drug has yet completed clinical trials and reached FDA approval. Competitor context: OpenAI released GPT-Rosalind in April 2026, and Google DeepMind operates Gemini for Science in the same research-tool category.

Disclosure: Claude, which generates this brief, is built by Anthropic.

President Trump signed EO 14409, "Promoting Advanced Artificial Intelligence Innovation and Security", on June 2, 2026, after an earlier 90-day pre-release draft was pulled in May over industry objections. The final order reduced the government's pre-release access window to 30 days, explicitly bars mandatory licensing or preclearance, and preserves developer freedom to release models. The order's first 30-day deadline directed Treasury, in consultation with the NSA and CISA, to form an AI cybersecurity clearinghouse in voluntary collaboration with the AI industry and critical infrastructure operators. That deadline fell July 2.

The order's 60-day deadline, due August 1, directs the Secretaries of Treasury, Defense, and Homeland Security to design the voluntary framework through which developers engage the federal government to have a model designated a "covered frontier model" and provide 30 days of government-only access. The classified benchmarking process to determine which models qualify has not been published. The EO explicitly states it cannot be used to authorize mandatory licensing or preclearance requirements. The FT separately reported July 3 that the White House is in advanced talks with OpenAI, Anthropic, and Google on a related but distinct voluntary announcement, possibly as soon as next week, which suggests the EO's own framework has not yet been finalized.

The order was catalyzed by Anthropic's Claude Mythos Preview model, which in April 2026 demonstrated superhuman vulnerability-finding capabilities that alarmed national security agencies. Commerce Secretary Howard Lutnick managed customer-by-customer access approvals for GPT-5.6 Sol manually during its limited preview, a sign that formal process did not exist at the time. Legal analysts at McDermott and A&O Shearman note that voluntary participation still creates competitive positioning effects, and that criteria for trusted-partner selection, who gets early access to pre-release frontier models, remain entirely unpublished.

On June 16, 2026, Palo Alto Networks Unit 42 researcher Ori Hadad published the "Pickle in the Middle" disclosure for a vulnerability in the Google Cloud Vertex AI Python SDK (google-cloud-aiplatform versions 1.139.0 and 1.140.0). The flaw's root cause: when a Vertex AI user uploads a model without specifying a custom staging bucket, the SDK constructs a bucket name using a deterministic pattern, project ID plus region, and checks only whether the bucket exists, not whether the victim owns it. An attacker with a standard Google Cloud account and knowledge of the victim's project ID (often public in GitHub config files or client-side JavaScript) could pre-create the expected bucket, wait for the victim's SDK to upload model artifacts to it, then replace the legitimate model file with a malicious pickle payload within an approximately 2.5-second race-condition window.

Once deployed, the malicious model executes the attacker's code inside Google's serving infrastructure. In Unit 42's proof-of-concept test environment, the stolen OAuth token accessed other model artifacts in the same tenant project, including a complete TensorFlow model with trained weights, as well as BigQuery metadata, access control lists, tenant logs, and internal container image paths. Google shipped an initial fix in v1.144.0 on March 31, 2026 (randomized bucket naming via UUID), and completed the remediation in v1.148.0 on April 15 (explicit bucket ownership verification in Model.upload()). Unit 42 reports no exploitation in the wild. This is the second predictable-bucket-name flaw in Vertex AI this year; CVE-2026-2473, patched in February, hit Vertex AI Experiments under a related attack path.

The practical takeaway for ML engineering teams: update google-cloud-aiplatform to v1.148.0 or later; explicitly define staging_bucket in all upload calls rather than relying on SDK defaults; and audit notebooks, CI pipelines, and training workflows, not only production services, for the SDK version in use. No CVE was assigned to the "Pickle in the Middle" bug at time of publication; this is the second consecutive Vertex AI bucket-squatting flaw without a CVE number, which means automated vulnerability scanners will not flag it.

The Financial Times reported on July 3 that the White House is in advanced talks with OpenAI, Google, and Anthropic on a voluntary framework to govern how the next tier of frontier AI models gets released, with an announcement possible as soon as the coming week. Per the FT, the framework would set benchmarks for advanced models, establish timelines, and clarify access rules for domestic and foreign users. Neither the White House, Anthropic, nor OpenAI responded to comment requests at time of FT publication. The Hill reported separately that the government's requests for model delays have spurred confusion among technology policy analysts, with Dean Ball, co-author of the Trump administration's AI Action Plan, writing that US federal AI policy had shifted "from implausibly libertarian to increasingly draconian and opaque" within weeks.

The context: Trump's June 2 EO 14409 directed agencies to design a voluntary pre-release framework by August 1, 2026, a deadline now 27 days away. Commerce Secretary Lutnick personally approved roughly 20 OpenAI GPT-5.6 Sol customers during the model's limited June preview, functioning as a de facto licensing office. Anthropic's Fable 5 and Mythos 5 were subject to a binding Commerce Department export-control directive, a legally distinct instrument from the voluntary EO framework, before being restored July 1 after Anthropic agreed to 99%-plus jailbreak filters, the first known technical condition attached to a US model reinstatement. AI Weekly's analysis notes that this arrangement operates as "a backdoor licensing regime built on existing Commerce Department authority, with conditions that shift without public notice."

The FT story is single-source, confirmed by wire pickup; the White House, Anthropic, and OpenAI did not respond, and no published benchmark list, criteria for foreign commercial customer access, or formal timeline has emerged as of this edition. Google is separately in talks with the government over advanced coding models with stronger cyber capabilities than prior versions, per sources familiar to TipRanks.

Section 2(d) of EO 14409, signed June 2, 2026, directed the Secretary of the Treasury, in consultation with the NSA and CISA, to form an "AI cybersecurity clearinghouse" within 30 days. The clearinghouse was to coordinate and deconflict scanning for software vulnerabilities, discover and validate them, and prioritize remediation and patch distribution across the AI industry and critical infrastructure operators. That 30-day deadline expired July 2. No Treasury announcement, Federal Register notice, or named clearinghouse has appeared publicly as of this edition.

The EO's second tranche of deadlines falls August 1, 2026, 27 days from today. By then, the Secretaries of Treasury, Defense, and Homeland Security are to have designed the voluntary "covered frontier model" engagement framework, including classified benchmarks that determine which models trigger the 30-day pre-release government access window. Those classified benchmarks have not been published; no draft criteria, no named agency lead for the benchmarking process, and no public engagement channel for developers have emerged. Legal analysts at McDermott note that until the classification criteria are published, developers cannot determine whether their in-development models will qualify, making the voluntary "opt-in" largely theoretical.

The significance for AI operators: the AISI (AI Safety Institute), CISA, and Treasury are the agencies whose guidance would shape how critical infrastructure customers should think about AI vulnerability disclosure. Without the clearinghouse, AI-discovered CVEs, a category growing at 46% above forecast this year per FIRST's revised 2026 estimate, flow through existing, fragmented disclosure channels with no AI-specific coordination layer. The FT's July 3 reporting suggests the White House is now trying to formalize this through a separate bilateral voluntary agreement with the labs, which may be intended to compensate for the EO's implementation lag.