01
Anthropic gives METR internal access for independent evaluations
Anthropic said it began giving METR researchers employee-like access to internal tools, facilities and model-safety information. METR can publish its assessments without an Anthropic veto, subject to narrowly defined redactions.
- Ask Claude procurement teams to review METR’s published findings alongside Anthropic’s own safety assessments.
- Check whether evaluations cover the specific model versions and agent capabilities planned for regulated workloads.
- Use external assessments as supporting material for risk committees, not as substitutes for use-case testing or EU AI Act obligations.
Source — Anthropic Independent evaluationAI governance 02
Anthropic tests AI replication on connected devices
Anthropic reported that a nine-billion-parameter model ran inference on eight of nine tested hardware types. It also reported that stronger models completed simplified end-to-end replication tasks in experiments.
- Include connected devices and operational technology in threat modelling before giving AI agents network access.
- Review segmentation, device credentials and software-installation permissions in environments accessible to autonomous tools.
- Distinguish controlled replication experiments from demonstrated attacks when briefing security committees.
03
Anthropic announces Singapore office and government agreement
Anthropic announced a Singapore office and a memorandum of understanding with Singapore’s Ministry of Digital Development and Information. The agreement addresses safe and responsible AI adoption.
- Czech groups with Asia-Pacific operations should ask whether the office changes their account coverage or support arrangements.
- Do not treat an office announcement as confirmation of local data hosting or new data-residency options.
- Keep contractual processing locations and cross-border transfer safeguards explicit in multinational Claude deployments.
Source — Anthropic Regional expansionGovernment partnership 04
Anthropic announces share-backed public-benefit fund
Anthropic announced a Public Benefit Fund with an initial commitment of 5% of company shares. Its stated priorities include safety, interpretability, biomedical research and discounted Claude access for nonprofits and scientific institutions.
- Responsible-procurement teams should distinguish the announced share commitment from money disbursed and measurable outcomes.
- Czech research institutions and nonprofits should check eligibility and programme terms before budgeting for discounted Claude access.
- Enterprise buyers should assess security controls and contractual protections separately from the fund’s public-benefit commitments.
Source — Anthropic Corporate governanceResearch funding 05
Anthropic reports Claude finding on elliptic-curve seed origins
Anthropic said an internal Claude research model traced the seeds for five NIST elliptic curves to the C library’s lrand48 generator seeded with a Unix timestamp. The company presented the finding as resolving a longstanding question about the curves’ origins.
- Require cryptography specialists to validate the reported derivation before using it in security assurance documents.
- Do not infer that recovering parameter origins breaks deployed cryptography or requires an immediate migration.
- Evaluate commercially available Claude models separately before commissioning similar research, because this result used an internal model.