Mythos enters security tests
Anthropic’s new model 'Mythos' is being trialled by major tech firms as part of Project Glasswing to hunt cybersecurity flaws before public release. Britain’s Bank of England has also joined Mythos testing under AI Safety Institute processes, according to recent coverage. (x.com) (meyka.com)
Anthropic has started letting major technology and infrastructure groups test Claude Mythos Preview, a new model built to find serious software flaws before wider release. (anthropic.com) Anthropic said Project Glasswing launched with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. The company also said more than 40 additional organizations that build or maintain critical software infrastructure received access to use the model on first-party and open-source systems. (anthropic.com) The basic idea is simple: defenders want the same kind of fast vulnerability hunting that attackers would want. Anthropic said Mythos Preview can identify and exploit previously unknown “zero-day” flaws across major operating systems and web browsers, so the company is keeping the model gated instead of releasing it publicly. (anthropic.com) Anthropic’s system card, published in early April 2026, called Mythos Preview its most capable frontier model to date and said it showed a large jump over Claude Opus 4.6 on several benchmarks. The company said the model is not planned for general availability in its current form. (anthropic.com 1) (anthropic.com 2) Britain’s AI Safety Institute, a government body created to study advanced model risks, has already worked with Anthropic on other safety research, including an October 2025 study on data poisoning with the Alan Turing Institute. The institute says its mission is to reduce surprise from rapid advances in advanced artificial intelligence. (aisi.gov.uk) (gov.uk) The Bank of England has been warning for months that wider use of artificial intelligence could create financial-stability risks if many firms rely on the same models or if hidden failures spread through the system. In a May 2024 speech, Financial Policy Committee member Jonathan Hall said artificial intelligence could amplify risk through herding and through dependence on a small number of providers. (bankofengland.co.uk) The central bank and the Prudential Regulation Authority also said in an April 2024 update that they were building their own capacity to supervise artificial intelligence in finance, including work on model risk, governance and operational resilience. That makes security testing of powerful models directly relevant to banks, payment systems and market infrastructure. (bankofengland.co.uk) Outside coverage has described the Bank of England as part of recent Mythos testing under AI Safety Institute processes, but Anthropic’s public Project Glasswing partner list does not name the bank. Anthropic’s own materials identify the launch partners and say the program also includes additional critical-software organizations without listing them all. (meyka.com) (anthropic.com) The immediate next step is closed testing, not a public rollout. Anthropic said it plans to use what partners learn in Project Glasswing to harden critical software and share lessons with the wider industry. (anthropic.com)
