Anthropic announced on April 7 that its new artificial intelligence model, the «Mythos», is so powerful that it cannot be released freely. Company head Dario Amodei is concerned about his skills in finding security gaps in software. Mythos has already identified critical vulnerabilitys in all major operating systems and browsers, including one that remained hidden for 27 years. Anthropic, who saw her annual revenue being launched at $30 billion on April 6, began the «Project Glasswing» to help companies strengthen their cybersecurity before the wide circulation of the model.
Analyticalally:
When in 2019 OpenAI completed training of a new major language model, GPT-2, the artificial intelligence laboratory initially stated that the tool was too dangerous to be released. Dario Amodei, then director of OpenAI research, insisted that the world needed adjustment time.
Eventually, the model was released later that same year. Since then a number of much stronger systems have developed, without Armageddon. However, seven years later, Mr Amodei, now head of Anthropic —the main competitor of OpenAI— He's worried again. On April 7 he stated that the last addition to the Claude model family, named «Mythos», it is so strong that it is still not allowed to be widely available. This time, maybe he's right.
Mythos has already identified critical security gaps in «each large operating system and browser», including one remaining unnoticed for 27 consecutive years.
According to Anthropic, Mythos' potential lies «significantly beyond anything we've trained in the past». The laboratory states particularly alarmed by the system's ability to detect software vulnerabilities and either repair them (operated defensively) or exploit them (acting as a hacker).
Such claims should normally be treated with scepticism. Anthropic built the model, conducted the tests — and apparently benefits from the impression that her system is smarter than any predecessor. The company is going through a period of impressive growth: on April 6, it announced that its annual revenue was launched at $30 billion, from just $9 billion at the end of last year. It is reasonable, then, to seek to maintain this dynamic.
However, there are good reasons to take its warnings into account. The first is their gravity: Anthropic argues that Mythos has already identified critical security gaps in «each large operating system and browser», including one remaining unnoticed for 27 consecutive years.
The second reason is the reaction of the market. Along with the suspension of traffic, Anthropic announced «Project Glasswing», an initiative that helps companies to shield their cyber defence through Mythos before it is widely available. The participation of colossae —such as Apple, Linux Foundation and CrowdStrike, and Google, which is a direct competitor— suggests that the threat is valid.
Mr Amodei's risk reduction strategy is considered rational. By providing priority, companies can check their code for weaknesses prior to its publication. However, Anthropic has a clear financial benefit from Project Glasswing: although it will cover the first $100 million of the cost of the initiative, in future it intends to charge participants five times more for the use of Mythos compared to the earlier model, Opus.
It is perhaps a price worth paying. Anthropic competitors are sure to develop sooner or later models with similar penetration skills. While workshops such as OpenAI and Google feature structured security policies, open source laboratories —Especially those in China— They tend to place less emphasis on shielding their systems.
Hackers are not the only ones who may be disturbed by Project Glasswing. The American government traditionally seeks to exploit weaknesses in its opponents' systems, «saving» unknown safety gaps (zero days) to use at the appropriate time. If Project Glasswing succeeds, it could disable much of the U.S. cybership.
Such a development would undoubtedly anger Pete Hegseth, US Secretary of Defense, who described Anthropic «risk to the supply chain» after a heated dispute with the Pentagon over the military use of its models. Even though a judge temporarily blocked this guy «orwellian» Mr Amodei seems to be going to continue to be «thorn» for the leadership of the ministry.
With data from Economist
