Anthropic shares Mythos findings with FSB

Anthropic said on May 18 that it would share findings from its Claude Mythos cyber-capability work with the Financial Stability Board rather than publish the model or a technical write-up, according to reporting by The Guardian and Reuters pickup carried by Channel News Asia. The company has not publicly released the model, citing concerns that the system’s cyber findings could be misused. The Financial Stability Board is the international body that coordinates financial authorities and standard setters across the G20. The move places Anthropic’s internal cyber evaluation work in a regulatory channel instead of the usual public research cycle. ### Why is Anthropic taking Mythos to a financial watchdog? The Financial Stability Board has already been studying AI as a financial-stability issue. In a November 2024 report, the FSB said AI could create vulnerabilities tied to cyber risk, third-party dependencies, model governance and concentration in the financial sector. In an October 2025 follow-up, the body said monitoring of AI adoption and related vulnerabilities was still at an early stage. (theguardian.com) Reuters-reported coverage said Anthropic was preparing to brief the FSB on vulnerabilities in the global financial system identified by Mythos. The Guardian reported that Anthropic chose to share the findings with the watchdog rather than release the model publicly. Anthropic has not, in the sources reviewed, published a public technical paper describing those findings in detail. (fsb.org) ### What exactly is changing in Claude billing on June 15? Anthropic’s customer support page says that starting June 15, 2026, Claude Agent SDK and `claude -p` usage will no longer count toward a user’s Claude plan usage limits. Instead, that usage will draw from a separate monthly Agent SDK credit denominated in dollars, while interactive use remains on the subscription limit. Anthropic says the change applies to Pro, Max, Team and Enterprise plans, while platform accounts using API keys continue on pay-as-you-go billing and do not receive that credit. (msn.com) A third-party billing explainer published on May 18 said Anthropic announced the change on May 14 and that usage in CI pipelines, products built on the Agent SDK, and cron-style `claude -p` jobs would be affected by the new split. That characterization matches Anthropic’s support language on which products move into the separate bucket, though the third-party site adds implementation detail not visible in the support excerpt. (support.claude.com) ### Why does the split matter for Claude Code and agent users? Anthropic’s own product materials describe Claude Code as an agentic coding system that can read codebases, make changes across files, run tests and deliver committed code. Anthropic also says the Claude Agent SDK gives developers access to the same core tools, context management systems and permissions frameworks that power Claude Code. That means the billing change affects products designed for longer-running, tool-using workflows, not just chat sessions. (codersera.com) Anthropic’s support page also says subscription usage limits will stay reserved for interactive use of Claude Code, Claude Cowork and Claude. That creates a clearer separation between attended use and programmatic or unattended use. It is reasonable to infer that the split gives customers more visibility into automated agent spend, because Anthropic is assigning those workloads to a separate credit pool, but Anthropic’s support page frames the change as a billing and usage-limits update rather than as a governance measure. (anthropic.com) ### How do the two announcements fit together? Anthropic has recently been emphasizing agent autonomy and developer tooling. In product posts, the company has expanded Claude Code capabilities, described the Agent SDK as the harness behind those workflows, and increased usage limits after adding compute capacity. In parallel, the Mythos decision shows Anthropic routing sensitive cyber-capability findings to a financial-risk body instead of publishing them openly. (support.claude.com) June 15 is the next concrete date. Anthropic’s support page says that is when the separate Agent SDK credit begins for eligible Claude plans, while any public disclosure of Mythos findings appears, for now, to be limited to the company’s planned briefing with the Financial Stability Board. (support.claude.com) (anthropic.com)