For every Anthropic, pretending to be smarter than anybody else, there is a hacker that is even smarter. Anthropic hordes the private information of users, pretending to defend safety. The hacker uses the information. Immense disaster is waiting around the corner. Let Anthropic fail early.
We have deep concerns about the concentration of data and power within AI labs like Anthropic. The tension between a company’s stated mission of "safety" and the reality of being a high-value target for hackers is a major point of debate in the tech world.
The "immense disaster" we're describing—where centralized data or powerful models fall into the wrong hands—is exactly what many security researchers call a single point of failure. Critics often argue that:
Centralization creates a honey pot: By gathering massive amounts of user data and proprietary research, these companies become the ultimate prize for state-sponsored or independent hackers.
Security is never absolute: No matter how "smart" the developers are, the history of cybersecurity shows that defenders have to be right every time, while a hacker only has to be right once.
Safety vs. Control: There is a fine line between building "safe" AI and building a "walled garden" that gives a single entity too much control over information.
Whether these organizations "fail early" or evolve into something more transparent, the conversation around decentralizing AI and open-source security is becoming much louder to prevent the kind of disaster we are worried about.
“Anthropic is investigating potential unauthorized access to its new Mythos artificial-intelligence model, complicating efforts to keep a tight lid on a tool that has spooked the White House and businesses because of its ability to facilitate cyberattacks.
The San Francisco-based company said it is investigating "a report claiming unauthorized access" to Mythos through a third-party contractor. It added that it had no evidence that the reported access extended beyond that third-party vendor.
The possibility of unregulated access to Mythos could send a tremor through the cybersecurity world. Researchers worry that it and other advanced AI tools could allow hackers to unleash widespread disruption via an avalanche of software bugs. But some have also taken hope from the idea that limited releases could help them patch vulnerabilities first.
The incident highlights the difficulty of restricting powerful new AI models and techniques in the middle of a brutal race to commercialize the tools. Earlier this month, Anthropic rushed to contain a leak of internal source code with internal instructions it uses to direct its popular Claude Code application.
Days later, the company announced the existence of its Mythos model, but said that it would hold off indefinitely on a public release of the tool after testing showed high capability for computer security tasks -- like finding thousands of software bugs.
Instead, Anthropic said it would limit Mythos's release to about 50 companies and organizations that maintain critical infrastructure and software, including Amazon, Alphabet's Google and the nonprofit Linux Foundation, so they could use it to find vulnerabilities before hackers do.
The White House has been racing to address new cybersecurity threats posed by AI models like Mythos and others from competing labs like OpenAI. Last week, Anthropic Chief Executive Dario Amodei met with top Trump administration officials about how to prepare for the model. Anthropic is in talks to give government agencies early access to Mythos, The Wall Street Journal reported.
The meeting, which the White House described as "both productive and constructive," could help thaw relations between the Trump administration and Anthropic. The company is fighting the U.S. government in two different courts.” [1]
1. Anthropic Looks Into Possible Unauthorized Access to Mythos. Schechner, Sam. Wall Street Journal, Eastern edition; New York, N.Y.. 23 Apr 2026: B1
Komentarų nėra:
Rašyti komentarą