Anthropic has on Wednesday launched Claude Mythos Preview, a brand new cyber AI mannequin, however the public can not use it.
Talking by way of a weblog publish, the corporate mentioned, “AI fashions have reached a degree of coding functionality the place they will surpass all however probably the most expert people at discovering and exploiting software program vulnerabilities.”
The AI firm mentioned it has estimated international cybercrime prices at round $500 billion a 12 months.
In accordance with Anthropic, its launch group for Mythos Preview contains Amazon Internet Companies, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Basis, Microsoft, NVIDIA, and Palo Alto Networks.
Greater than 40 different organizations that construct or preserve vital software program additionally obtained entry. Anthropic mentioned it’s going to present as much as $100 million in utilization credit and $4 million in direct assist for open supply safety teams.
In its press launch, Anthropic claims that Mythos Preview has allegedly discovered hundreds of high-severity vulnerabilities throughout each main working system and each main net browser.
One instance was a 27-year-old flaw in OpenBSD that might let an attacker remotely crash a machine simply by connecting to it. One other was a 16-year-old flaw in FFmpeg hiding in code that automated instruments had hit 5 million occasions with out catching the difficulty.
The mannequin additionally discovered and chained a number of flaws within the Linux kernel so an attacker might transfer from bizarre consumer entry to full management of a machine.
Anthropic mentioned for different bugs, it plans to publish cryptographic hashes now and can reveal extra as soon as fixes are in place, because the mannequin discovered practically all of these vulnerabilities and constructed many associated exploits by itself.
On CyberGym, Mythos Preview scored 83.1% on vulnerability copy, in contrast with 66.6% for Claude Opus 4.6. VentureBeat individually reported 93.9% on SWE-bench Verified, versus 80.8% for Opus 4.6.
Anthropic then defined that latest frontier techniques have minimize the price, effort, and ability wanted to search out and exploit safety holes.
Glasswing provides companions a head begin in a quicker cyber battle
Beneath Undertaking Glasswing, companions will use Mythos Preview for defensive work on inner techniques and open supply code.
Anthropic mentioned the work will embody native vulnerability detection, black field testing of binaries, endpoint safety, and penetration testing.
After the analysis preview, individuals will be capable to entry the mannequin by the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at $25 per million enter tokens and $125 per million output tokens.
The corporate additionally mentioned it gave $2.5 million to Alpha-Omega and OpenSSF by the Linux Basis, plus $1.5 million to the Apache Software program Basis.
AWS mentioned it analyzes greater than 400 trillion community flows a day, Microsoft mentioned the mannequin confirmed features on CTI-REALM, CrowdStrike mentioned the hole between discovering a flaw and exploiting it has collapsed, and Google mentioned it’s going to make the mannequin accessible by Vertex AI, whereas Palo Alto Networks mentioned defenders want these instruments earlier than attackers get them.
The New York Occasions reported that late final 12 months, Anthropic mentioned state-backed Chinese language hackers used its AI in an effort to focus on about 30 firms and authorities businesses, with human operators doing solely 10% to twenty% of the work.
The report additionally mentioned attackers are already utilizing AI to draft phishing emails, write ransom notes, kind stolen knowledge, and velocity up breach gross sales.
Your financial institution is utilizing your cash. You’re getting the scraps. Watch our free video on changing into your personal financial institution