Guardian

Anthropic investigates report of rogue access to hack-enabling Mythos AI

BBC News

Claude Mythos AI unauthorised access claim probed by Anthropic

Financial Times

Anthropic investigating unauthorised access of powerful Mythos AI model

Bloomberg

Anthropic Mythos Model Is Being Accessed By Unauthorized Users

Anthropic

Project Glasswing: Securing critical software for the AI era

MarioNawfal

ChrisRMcGuire

The Register

Anthropic Mythos shaping up as nothingburger

RTE.ie

Powell warned bank CEOs on Anthropic model risks: sources

The Telegraph

Bank of England raises alarm over threat from AI ‘too dangerous to release’

Reuters

Japan's financial watchdog to meet top banks to discuss Mythos AI model

CTVNews

Canadian banks, regulators discussed Mythos AI, minister to meet with Anthropic

Metaculus

AI Unauthorized Access Before 2033?

The AI company Anthropic has confirmed it is investigating a report of unauthorized access to its Claude Mythos Preview model via a third-party vendor environment, telling multiple media outlets that it has found no evidence the breach extended beyond that vendor's systems.

The AI company Anthropic has confirmed it is investigating a report of unauthorized access to its Claude Mythos Preview model via a third-party vendor environment, telling multiple media outlets that it has found no evidence the breach extended beyond that vendor's systems.

According to Bloomberg, a handful of users in a "private online forum" gained access to Mythos the same day as Anthropic announced a limited release to companies including Apple, Amazon, Cisco and Goldman Sachs for testing purposes under Project Glasswing.

According to Bloomberg, a handful of users in a "private online forum" gained access to Mythos the same day as Anthropic announced a limited release to companies including Apple, Amazon, Cisco and Goldman Sachs for testing purposes under Project Glasswing.

The group, which is part of a private Discord channel dedicated to finding unreleased AI models, allegedly employed tools commonly used by cybersecurity researchers as well as knowledge gleaned from the <a target="_blank" rel="noopener noreferrer nofollow" href="https://www.verity.news/story/2026/lapsus-hits-mercor-via-litellm-supply-chain-attack-2">Mercor breach</a> about Anthropic's AI formatting to locate the model.

The group, which is part of a private Discord channel dedicated to finding unreleased AI models, allegedly employed tools commonly used by cybersecurity researchers as well as knowledge gleaned from the Mercor breach about Anthropic's AI formatting to locate the model.

A person familiar with the matter also told the publication that one member of the Discord group had existing access to Anthropic's systems and models as a contractor for evaluating the company's technologies, which they used to access Mythos.

A person familiar with the matter also told the publication that one member of the Discord group had existing access to Anthropic's systems and models as a contractor for evaluating the company's technologies, which they used to access Mythos.

Mythos, Anthropic claims, is capable of identifying and exploiting vulnerabilities across every major operating system or web browser, having already discovered thousands of zero-day vulnerabilities. Despite this, the source stated that the group had not used the model for cybersecurity purposes, instead constructing "simple websites" with the tool to avoid detection from Anthropic.

Mythos, Anthropic claims, is capable of identifying and exploiting vulnerabilities across every major operating system or web browser, having already discovered thousands of zero-day vulnerabilities. Despite this, the source stated that the group had not used the model for cybersecurity purposes, instead constructing "simple websites" with the tool to avoid detection from Anthropic.

Anthropic's limited release of Mythos has prompted an international reaction, with central banks and government officials holding talks with industry leaders in the <a target="_blank" rel="noopener noreferrer nofollow" href="https://www.verity.news/story/2026/us-regulators-banks-meet-over-anthropic-ai-cyber-risks">U.S.</a>, U.K. and Canada to discuss the potential cybersecurity risks posed by AI. Japan, meanwhile, intends to hold its own such meeting on Friday.

Anthropic's limited release of Mythos has prompted an international reaction, with central banks and government officials holding talks with industry leaders in the U.S., U.K. and Canada to discuss the potential cybersecurity risks posed by AI. Japan, meanwhile, intends to hold its own such meeting on Friday.

Guardian #

BBC News #

Financial Times #

Bloomberg #

Anthropic #

The Register #

RTE.ie #

The Telegraph #

Reuters #

CTVNews #

Metaculus #

Anthropic confirmed it is investigating a report of unauthorized access to its Claude Mythos Preview model through a third-party vendor environment. The company said it has no evidence the access extended beyond that vendor's systems.

Wall Street Journal

Anthropic confirmed it is investigating a report of unauthorized access to its Claude Mythos Preview model through a third-party vendor environment. The company said it has no evidence the access extended beyond that vendor's systems.

According to Bloomberg, a handful of users in a private online forum gained access to Mythos on the day Anthropic announced a limited release to companies including Apple, Amazon, Cisco and Goldman Sachs for testing purposes.

According to Bloomberg, a handful of users in a private online forum gained access to Mythos on the day Anthropic announced a limited release to companies including Apple, Amazon, Cisco and Goldman Sachs for testing purposes.

One person in the group had existing permissions as a contractor for Anthropic, which they used to access the model. The group also employed internet sleuthing tools commonly used by cybersecurity researchers to locate the model.

One person in the group had existing permissions as a contractor for Anthropic, which they used to access the model. The group also employed internet sleuthing tools commonly used by cybersecurity researchers to locate the model.

Anthropic developed Mythos Preview under its Project Glasswing initiative, restricting access to roughly 50 organizations maintaining critical infrastructure, including Amazon, Google, Microsoft and the Linux Foundation.

Anthropic developed Mythos Preview under its Project Glasswing initiative, restricting access to roughly 50 organizations maintaining critical infrastructure, including Amazon, Google, Microsoft and the Linux Foundation.

The UK's AI Security Institute assessed that Mythos was a step up from prior models in cyber threat capability, noting it could carry out multi-step attacks and discover IT system weaknesses without human intervention.

The UK's AI Security Institute assessed that Mythos was a step up from prior models in cyber threat capability, noting it could carry out multi-step attacks and discover IT system weaknesses without human intervention.

Anthropic committed $100 million in model usage credits to Project Glasswing and donated $4 million to open-source foundations, including Alpha-Omega, OpenSSF and the Apache Software Foundation, to help maintainers address emerging security risks.

Anthropic committed $100 million in model usage credits to Project Glasswing and donated $4 million to open-source foundations, including Alpha-Omega, OpenSSF and the Apache Software Foundation, to help maintainers address emerging security risks.

Anthropic's Claude Mythos was breached on its launch day by a Discord group that guessed its URL — and that should terrify everyone. The Pentagon flagged Anthropic as a supply chain risk, and global regulators are already scrambling. If a casual group of hobbyists can crack a restricted cybersecurity AI in 24 hours, the case for stricter export controls and security mandates on frontier AI models is airtight.

Anthropic's Claude Mythos was breached on its launch day by a Discord group that guessed its URL — and that should terrify everyone. The Pentagon flagged Anthropic as a supply chain risk, and global regulators are already scrambling. If a casual group of hobbyists can crack a restricted cybersecurity AI in 24 hours, the case for stricter export controls and security mandates on frontier AI models is airtight.

Chris McGuire

Mario Nawfal

The Mythos breach sounds alarming but the actual damage is basically zero — no Anthropic systems were compromised, and the model itself is nowhere near as dangerous as advertised. Early testing from Mozilla and AWS shows Mythos hasn't surpassed elite human researchers, and independent analysis found Anthropic's vulnerability claims were heavily exaggerated. Attackers already have cheaper tools that do the same job, making the hysteria around this breach wildly overblown.

The Mythos breach sounds alarming but the actual damage is basically zero — no Anthropic systems were compromised, and the model itself is nowhere near as dangerous as advertised. Early testing from Mozilla and AWS shows Mythos hasn't surpassed elite human researchers, and independent analysis found Anthropic's vulnerability claims were heavily exaggerated. Attackers already have cheaper tools that do the same job, making the hysteria around this breach wildly overblown.

Artificial Intelligence

Anthropic Probes Unauthorized Access to Claude Mythos AI

Cybercrime

The fact that Mythos — supposedly the most powerful AI tool ever developed — was breached on its launch day by a Discord group that effectively guessed its URL should terrify everyone. If a casual group of hobbyists can crack a restricted cybersecurity AI within hours, the case for stricter export controls and security mandates for frontier AI models is airtight.

The fact that Mythos — supposedly the most powerful AI tool ever developed — was breached on its launch day by a Discord group that effectively guessed its URL should terrify everyone. If a casual group of hobbyists can crack a restricted cybersecurity AI within hours, the case for stricter export controls and security mandates for frontier AI models is airtight.

The Mythos breach sounds alarming, but the actual damage is basically zero. No Anthropic systems were compromised, and, as Mozilla and AWS's testing has proven, the model itself is nowhere near as dangerous as Anthropic advertised. Given that attackers already have cheaper tools that do the same job, the hysteria surrounding this breach is wildly overblown.

The Mythos breach sounds alarming, but the actual damage is basically zero. No Anthropic systems were compromised, and, as Mozilla and AWS's testing has proven, the model itself is nowhere near as dangerous as Anthropic advertised. Given that attackers already have cheaper tools that do the same job, the hysteria surrounding this breach is wildly overblown.

There is a 99% chance that an AI system will be reported to have independently gained unauthorized access to another computer system before 2033, according to the Metaculus prediction community.

There is a 99% chance that an AI system will be reported to have independently gained unauthorized access to another computer system before 2033, according to the Metaculus prediction community.

Anthropic Probes Unauthorized Access to Claude Mythos AI

Anthropic Probes Unauthorized Access to Claude Mythos AI

The Spin

Metaculus Prediction

Go Deeper

Sign Up for Our Free Newsletters
Sign Up for Our Free Newsletters

Anthropic Probes Unauthorized Access to Claude Mythos AI

Anthropic Probes Unauthorized Access to Claude Mythos AI

The Spin

Metaculus Prediction

Go Deeper

Sign Up for Our Free NewslettersSign Up for Our Free Newsletters

Sign Up for Our Free Newsletters
Sign Up for Our Free Newsletters