Anthropic Limits Claude Mythos AI Over Hacking Concerns
Synopsis
In a landmark move for AI safety, Anthropic has limited access to its Claude Mythos model due to its unprecedented ability to uncover software vulnerabilities autonomously. The model's discovery of a long-standing Linux kernel bug has sparked urgent meetings between the U.S. Treasury, U.K. regulators, and major bank CEOs. While Anthropic is moving cautiously, industry experts warn that the era of AI-driven hacking is here, requiring a total rethink of how critical infrastructure and financial systems are protected. With similar models expected to go public within the next 18 months, the pressure is on for global institutions to harden their software against machine-speed attacks.
While training its latest AI model, Claude Mythos, Anthropic severely limited access after it found that the software could independently discover and exploit serious vulnerabilities. The move has sparked national security conversations between government regulators and America’s biggest financial institutions.
Key Highlights
- Claude Mythos autonomously exploits zero-day flaws and, once an uncovered bug in the Linux kernel.
- Very limited access to C-I-C Security Researchers only
- Potential penalties against companies for not fortifying AI-vulnerable software due to new EU Cyber Resilience Act (CRA) obligations.
Major flaws lead to Mythos being limited by Anthropic
Anthropic said in a statement released Wednesday that it had limited access to its new Mythos model of AI to a handful of partners with the highest levels of security clearance after testing revealed it is able to hack itself into autonomous operations faster than previously imagined. The model showed its potential in uncovering layers of software vulnerabilities, including a flaw in the Linux kernel that had gone undetected for nearly three decades, with absolutely no human guidance.
Jack Clark, an Anthropic co-founder, told the that the company opted for a slow rollout because even if a system can identify and potentially exploit weaknesses at machine speed, clicking or typing takes more time. The business is now partnering with external experts to announce and repair these vulnerabilities earlier than the criminals can take advantage of them.
Global regulators scramble to gauge systemic risk
The surprise has rocked the world banking industry, leading to emergency meetings of top bank executives with the U.S. Treasury Secretary and British regulators. It is feared that an adversarial AI may one day be deployed to trigger the automation of zero-day vulnerabilities within not only trading systems, but payment networks and customer platforms which could lead to systemic market instability. At this point, specialists including Nik Kairinos from RAIDS AI argue that the debate is no longer one of theory but an active scramble as institutions catch on that actual frontier AI models can dismantle essential structures in real time.
Cybersecurity experts urge industrialisation of cybersecurity
Security experts describe the limited launch of Mythos as simply buying time, adding that similar programs from international rivals are bound to follow. Ansgar Dodt, from Thales, believes that businesses have to operate on the presumption that their software is under constant stress testing and reverse engineering by AI tools. Developers are being advised to adopt security by design, steps which include, but are not limited to code encryption, logic obfuscation and embedding defences that catch tampering as it occurs, to endure this transition.
Companies that fail to make their applications resilient against these AI-driven threats will face harsh consequences under the new Cyber Resilience Act, including heavy fines, product recalls, and restrictions on access to markets.
FAQs
- What is Claude Mythos?
It is a new state-of-the-art AI model from Anthropic that accidentally found holes in computer software security vulnerabilities.
- Why is it considered dangerous?
It can hunt bugs that have gone unnoticed by humans for decades, making it applicable to hackers attempting rapid assaults on banks or power grids.
- How accessible is Mythos?
For this reason, the answer is no: Anthropic has limited it to a small number of researchers and government agencies only so that ill-intentioned people cannot abuse it.
- How are banks responding?
That is why, according to one bank CEO looking at government officials, CEOs of banks are assessing how they keep their trading and payment systems free from attacks powered by generative AI.
Follow Inspirepreneur Magazine for daily global business news.
u
At Inspirepreneurs Magazine, covering entrepreneurship, business failures, and the human stories behind the world's most ambitious founders. She writes at the intersection of strategy and storytelling.
You Might Also Like
EU Accuses Google and Apple, Triggering Trump Showdown
Oil prices surge nearly 20% as Iran war fuels supply fears
