The AI Brief

Anomaly Innovations, the company behind the open-source coding agent OpenCode, posted to X at 18:08 UTC on October 17 confirming the breach. The company said the incident was "related to a large scale distillation attack" and affected approximately 1.4% of OpenCode Console users who had linked external provider accounts. Anomaly Innovations said it emailed every affected user that morning and is working with Anthropic, OpenAI, and Google to remediate. The breach was first surfaced publicly by security researcher Matt Johansen roughly two and a half hours before OpenCode's confirmation.

The exposed credentials were third-party API keys stored inside Console workspaces on a legacy system the company was still decommissioning. Affected users reported unauthorized charges appearing on connected provider accounts within hours of the breach. OpenCode has not disclosed the total absolute number of affected accounts, the full technical path of the attack, or whether query logs from the attacker's inference sessions are recoverable. Users who ran OpenCode locally with environment-variable keys rather than saving credentials inside the Console dashboard appear to have lower exposure through this incident, though OpenCode has not confirmed a complete scope boundary.

The immediate action for any OpenCode Console user with linked provider keys is to revoke and replace those keys at the provider dashboard (OpenAI, Anthropic, or Google), update any applications or CI pipelines using them, and audit recent usage logs for anomalous spend spikes. The incident is the second significant AI coding-agent credential event of the year following the GitSpawn class of vulnerabilities disclosed in September 2026.

The 60-plus figure comes from Michel Leonard, chief economist at the Insurance Information Institute, cited in Global Finance Magazine's October 2026 issue. It refers to carrier groups that have filed to adopt the ISO generative AI exclusion endorsements (CG 40 47, CG 40 48, and CG 35 08) made available by Verisk's ISO on January 1, 2026. Berkshire Hathaway and Chubb have already received regulator approval to attach exclusion clauses to commercial liability policies. AIG, WR Berkley, and Great American have each filed regulatory paperwork to exclude AI-related claims from commercial coverage. Berkley's "absolute" AI exclusion reaches D&O, E&O, and fiduciary liability in a single endorsement.

The market dynamic follows a pattern familiar from early cyber insurance: carriers exit base coverage first, then specialists (primarily Lloyd's-backed managing general agents) sell affirmative AI cover as a separate, more expensive layer. Unlike cyber, where loss history now allows pricing, no comparable AI claims data set yet exists, which means specialty premiums carry substantial uncertainty loading. Industry officials are floating a federal liability backstop as a long-term solution, though no statutory proposal has cleared committee. The EU AI Liability Directive, the European parallel, was effectively killed by mid-2026, leaving both jurisdictions relying on market mechanisms to set de facto liability rules.

Anthropic published the experiment on October 16, 2026, tied to an arXiv preprint (2610.22367) submitted October 15 by Andi Gu, Alek Kemeny, Susanne Yelin, and Mikhail Lukin, affiliates of Anthropic's STEM Fellows Program, Harvard, and QuEra Computing. Claude Mythos 5 was given plain-language descriptions of eight standard computational primitives drawn from chemistry, optimization, physics simulation, and number theory. For each task, the model proposed a circuit written in Python, received a score from an independent automated evaluator, and iterated with no human input between rounds, running inside the open-source OpenEvolve search framework.

The model beat published baselines on seven of the eight tasks. The most widely cited single comparison is a modular inversion circuit that uses up to 17 times fewer gates than the prior best approach (with a 12% qubit penalty). Measured on the combined gates-times-qubits metric, improvements ranged from 4.1x to 39x across the six tasks where full cost data is reported. On the eighth task (elliptic curve modular inversion), Claude invented an architecture with no prior published equivalent. The paper is an unreviewed preprint; no hardware execution is described, and the results are classical simulation estimates. The Harvard collaborators Mikhail Lukin and Susanne Yelin were previously cited in aggregator coverage of the same work.

OpenAI announced on October 15 a $5 million contribution to a new Simons Foundation initiative. The full initiative totals $15 million, co-funded by OpenAI, Anthropic, and the Simons Foundation itself. The funds are designated for workshops, conferences, and special programs to help mathematicians examine, verify, and extend AI-generated results, with model and compute access included for Simons-supported researchers. The donation followed a public statement from the Association for Human Mathematics (AHM) that OpenAI showed "total disregard for the norms of scientific research" in publishing over 700 AI-generated manuscripts, including claims on the quasi-Riemann hypothesis and the Unique Games Conjecture, that OpenAI-led organization Sam Altman disclosed October 6.

The backlash has a specific edge: TechCrunch reported that two human mathematicians say they were close to a similar result on one of the claimed problems and question whether OpenAI's model drew on their unpublished work. One of those mathematicians, Levent Alpoge, works at Anthropic in a research capacity and was collaborating with the other in a private capacity. OpenAI denies having seen their work. None of the 722 manuscripts have been confirmed by external peer review. The Simons Foundation, which manages more than $4 billion in assets and has funded mathematical sciences since 1994, will administer and distribute the donated funds independently.

Anthropic CEO Dario Amodei and EverWind Fuels announced the lease on October 16, 2026, alongside Nova Scotia Premier Tim Houston. EverWind is headquartered in Halifax and is pursuing wind and green hydrogen projects in the province. The initial lease covers the first phase of EverWind's Point Tupper campus: 95 MW of capacity in an 18,600-square-meter facility targeted to come online in late 2027. The full campus build-out targets 420 MW by 2028. Anthropic and EverWind have not disclosed the financial terms of the lease itself.

The broader EverWind project carries a cost above C$10 billion, described as the largest privately funded investment in Nova Scotia's history. More than C$8 billion is earmarked for the data center and C$2 billion for onshore wind generation projects that will supply it. The campus is projected to create approximately 2,000 construction jobs and 215 permanent positions, with salaries described by EverWind in the C$135,000 to C$185,000 range. Anthropic's head of public policy for Canada, Jacob Glick, said the company committed to protecting local ratepayers and water resources. Anthropic noted that Canada is currently the sixth-largest user of Claude worldwide by traffic.