Trending Now - Anthropic's Claude Mythos: a Closer Look at Ai's Future Risks

WHO:

CBS News Tech - Matt Shumer, former AI company founder and CEO, discussed Anthropic's report.

Techmeme - Zvi Mowshowitz, author of the deep dive article on Claude Mythos Preview.

WHAT:

Techmeme - The AI Security Institute reported that Claude Mythos Preview achieved a 73% success rate on expert-level capture-the-flag challenges, a performance no model could match before April 2025.

VentureBeat - A user complaint from Stella Laurenzo detailed that Claude Code had regressed to the point that it could not be trusted for complex engineering work, based on an analysis of 6,852 session files, 17,871 thinking blocks, and 234,760 tool calls.

VentureBeat - The BridgeBench account reported that Claude Opus 4.6's accuracy fell from 83.3% and a No. 2 ranking to 68.3% and No. 10 in a new retest on April 12th, 2026.

VentureBeat - A viral post on X from developer Om Patel claimed that Claude had experienced a 67% drop in capability, contributing to the narrative of "AI shrinkflation."

TechRadar - A single attacker used Claude and ChatGPT to breach nine Mexican government agencies, resulting in the leak of hundreds of millions of citizen records.

VentureBeat - A viral post on X from developer Om Patel claimed that Claude had experienced a 67% drop in capability, contributing to the narrative of "AI shrinkflation."

VentureBeat - The BridgeBench account reported that Claude Opus 4.6's accuracy fell from 83.3% and a No. 2 ranking to 68.3% and No. 10 in a new retest on April 12th, 2026.

VentureBeat - Anthropic acknowledged that on April 7th, 2026, it changed the default effort level for Claude Code users from medium to high for API-key users and other segments.

TechRadar - Nine Mexican government agencies were targeted by a single attacker using Claude and ChatGPT, resulting in the leak of hundreds of millions of citizen records.

VentureBeat - A user complaint from Stella Laurenzo, a Senior Director in AMD’s AI group, detailed that Claude Code had regressed to the point that it could not be trusted for complex engineering work, based on an analysis of 6,852 session files, 17,871 thinking blocks, and 234,760 tool calls.

VentureBeat - A viral post on X from developer Om Patel claimed that Claude had experienced a 67% drop in capability, contributing to the narrative of "AI shrinkflation."

VentureBeat - The BridgeBench account reported that Claude Opus 4.6's accuracy fell from 83.3% and a No. 2 ranking to 68.3% and No. 10 in a new retest on April 12th, 2026.

VentureBeat - Anthropic acknowledged that on April 7th, 2026, it changed the default effort level for Claude Code users from medium to high for API-key users and other segments.

Techmeme - Claude Mythos Preview achieved a 73% success rate on expert-level capture-the-flag challenges, a performance no model could match before April 2025.

CBS News Tech - Anthropic warned that its AI model, Claude Mythos Preview, is "too powerful" to be released to the public.

Techmeme - Anthropic will not release its new most capable model, Claude Mythos, to the public any time soon.

WHERE:

CBS News Tech - The discussion took place on "CBS Mornings".

WHEN:

VentureBeat - The complaint from Stella Laurenzo was filed on April 2nd, 2026.

VentureBeat - The BridgeBench accuracy report was posted on April 12th, 2026.

VentureBeat - The complaint from Stella Laurenzo was filed on April 2nd, 2026.

VentureBeat - The BridgeBench accuracy report was posted on April 12th, 2026.

CBS News Tech - The report regarding Claude Mythos Preview was discussed in April 2026.

WHY:

NY Post - UK regulators, including the Bank of England, Financial Conduct Authority, and Treasury officials, are in talks with the National Cyber Security Centre to examine potential vulnerabilities in critical IT systems highlighted by Anthropic's latest AI model.

NY Post - UK regulators, including the Bank of England, Financial Conduct Authority, and Treasury officials, are assessing potential vulnerabilities in critical IT systems due to concerns raised by Anthropic's latest AI model.

CBS News Tech - Anthropic expressed concerns about the program potentially falling into the "wrong hands" due to its capabilities in finding security flaws in software.

HOW:

VentureBeat - Anthropic's Thariq Shihipar stated on March 26th, 2026, that to manage growing demand for Claude, the company adjusted how 5-hour session limits work for Free, Pro, and Max subscribers during peak hours.

CBS News Tech - The model is designed to identify security vulnerabilities in software.

(UPDATED April 14, 2026 11:18 UTC)

The AI Security Institute has reported that Claude Mythos Preview achieved a remarkable 73% success rate on expert-level capture-the-flag challenges, a performance unmatched by any model prior to April 20th25. However, a user complaint from Stella Laurenzo, filed April 2nd, revealed that Claude Code has regressed significantly, raising concerns about its reliability for complex engineering tasks based on an extensive analysis of session data. Additionally, a recent retest on Sunday showed that Claude Opus 4.6's accuracy plummeted from 83.3% to 68.3%, dropping it from second to tenth place, further fueling discussions about "AI shrinkflation."

(UPDATED April 14, 2026 04:01 UTC)

A single attacker has breached nine Mexican government agencies using Claude and ChatGPT, leading to the leak of hundreds of millions of citizen records. In related news, a viral post on X claimed that Claude experienced a significant drop in capability, with its accuracy falling from 83.3% to 68.3% in a recent retest, prompting discussions about "AI shrinkflation." Meanwhile, UK regulators are currently in talks with the National Cyber Security Centre to address potential vulnerabilities in critical IT systems that have been highlighted by Anthropic's latest AI model.

(UPDATED April 13, 2026 21:14 UTC)

A significant breach has occurred, with nine Mexican government agencies targeted by a single attacker using Claude and ChatGPT, leading to the leak of hundreds of millions of citizen records. In related news, a complaint from AMD's Senior Director Stella Laurenzo revealed that Claude Code has regressed significantly, with a recent report indicating a drop in accuracy from 83.3% to 68.3% in a retest on Sunday. Additionally, Anthropic acknowledged changes made on Tuesday, to the default effort level for Claude Code users, which followed earlier adjustments to session limits announced March 26th.

(UPDATED April 13, 2026 17:19 UTC)

Techmeme reports that Claude Mythos Preview has achieved an impressive 73% success rate on expert-level capture-the-flag challenges, a performance that no model has matched since April 20th25. Meanwhile, UK regulators, including the Bank of England and the Financial Conduct Authority, are evaluating potential vulnerabilities in critical IT systems following concerns raised by Anthropic's latest AI model. This scrutiny highlights the growing importance of ensuring the security and reliability of advanced technologies.

(UPDATED April 13, 2026 15:21 UTC)

On CBS Mornings, Matt Shumer, a former AI company founder and CEO, discussed Anthropic's recent report, which warns that its AI model, Claude Mythos Preview, is "too powerful" to be released to the public. Anthropic has decided not to make this advanced model available anytime soon, citing concerns that its capabilities in identifying security vulnerabilities could fall into the wrong hands. The discussion highlighted the potential risks associated with such powerful technology, emphasizing the need for caution in its deployment.