Anthropic Delays Release of Claude Mythos Amid Concerns Over Cybersecurity
SAN FRANCISCO, April 11 — Anthropic’s decision to postpone the unveiling of its latest AI model, Claude Mythos, which purportedly possesses extraordinary coding capabilities, has incited both apprehension and skepticism within the tech community.
As a notable player in the fiercely competitive landscape of artificial intelligence, Anthropic’s promotion of its technology not only bolsters its market presence but also enhances its allure amidst rumors of a looming public offering.
“The cybersecurity risks associated with Mythos must be taken with utmost seriousness,” remarked David Sacks, a prominent entrepreneur and investor who chairs President Donald Trump’s tech advisory council.
“Nevertheless, Anthropic has a documented history of utilizing fear-provoking strategies.”
The emergence of Mythos has birthed worries about the potential for hackers to deploy an army of AI agents capable of breaching digital defenses with alarming efficiency.
During this week’s HumanX AI conference in San Francisco, Alex Stamos of Corridor, a startup dedicated to AI safety, acknowledged the genuine threat posed by these agentic hackers.
Stamos also humorously critiqued what he termed Anthropic’s “marketing schtick.”
“They’ve created whimsical illustrations for products that are inherently perilous, yet they limit public access,” he pointedly noted, drawing a parallel to the hypothetical announcement of nuclear capabilities via a charming Calvin and Hobbes cartoon.
The leading executives of America’s major banking institutions convened recently with Federal Reserve Chairman Jerome Powell and Treasury Secretary Scott Bessent to deliberate on the security ramifications of the unreleased Claude Mythos, according to reports from Friday.
“The Mythos model signifies implications far more profound than merely another enhancement in artificial intelligence,” stated Shlomo Kramer, co-founder and CEO of Cato Networks, in a blog entry.
“It heralds a pivotal shift that could transform the dynamics between aggressors and defenders in the realm of cyberspace.”
A selectively disclosed preview of Mythos was shared with partner organizations this week, under an initiative dubbed Project Glasswing, encompassing collaborations with Amazon, Apple, Microsoft, Google, Cisco, CrowdStrike, and JPMorgan Chase.
According to Anthropic and its partners, Mythos can autonomously comb through extensive codebases to identify and interlink unknown security flaws in diverse software, from operating systems to web browsers.
Crucially, this process can be executed at a velocity and scale beyond human capacity, potentially incapacitating financial institutions, healthcare facilities, or critical national infrastructure within mere hours.
“Tasks that formerly necessitated elite specialists can now be accomplished by software agents,” Kramer emphasized.
“The ramifications will be a marked escalation in vulnerability discovery, a veritable tsunami of both known and unknown exploitations.”
‘Agent-to-Agent War’ Emerges
At HumanX, the prevailing sentiment was that AI agents, already proficient in coding, would excel in uncovering software vulnerabilities.
“We have entered an era where human coders are outmatched by superhuman AI models adept at identifying bugs,” Stamos argued.
He predicts a looming landscape wherein humans will oversee AI agents to safeguard networks from threats posed by hackers employing equivalent technologies for attacks.
Stamos referred to this dynamic as an “agent-to-agent war,” with humans relegated to advisory roles on the sidelines.
Wendy Whitmore of Palo Alto Networks anticipates “some form of catastrophic attack” this year attributable to AI agent capabilities.
“What disturbs me most is the approaching wave of new vulnerabilities that will be detected by AI,” cautioned Adam Meyers of CrowdStrike.
Meyers posited that incorporating a diminutive AI model directly into malevolent code targeting networks would be a strategic avenue likely to be pursued by hackers.
“The ultimate weapon would be adaptive malware devoid of preset programming,” Meyers concluded.
Source link: Malaymail.com.
