Anthropic Unveils New AI Models Amid Cybersecurity Concerns
In a strategic advancement, Anthropic has recently expanded its Project Glasswing, aimed at bolstering cybersecurity, by launching two new AI models: Claude Mythos 5 and Claude Fable 5. This move comes shortly after the project’s expansion, highlighting the company’s commitment to enhancing its capabilities in the realm of cybersecurity.
According to an announcement made on June 9, Claude Mythos 5 has been described by Anthropic as “an upgrade to Claude Mythos Preview.” In a bold claim, the company asserts that it possesses “the strongest cybersecurity capabilities of any model in the world.” This assertion is indicative of Anthropic’s ambition to lead the industry in AI-driven cybersecurity solutions.
The deployment of Mythos 5 will initially be facilitated through Project Glasswing, in a collaborative effort with the U.S. government. Looking ahead, the company envisions a broader availability of this advanced model through what it terms a “broader trusted access program.” This approach is aimed at ensuring responsible usage and accessibility in the future.
In tandem with Mythos 5, Anthropic has introduced Fable 5, which is based on the same foundational frontier AI model as Mythos 5. However, it includes additional safeguards, particularly concerning cybersecurity. The company has indicated that Fable 5 has been designed with the potential for misuse in mind, emphasizing the need for protective measures due to the serious consequences that could arise from malicious use.
One notable feature of these guardrails involves a system wherein queries related to specific topics made by Fable 5 will yield responses from Claude Opus 4.8, Anthropic’s subsequent model that is widely accessible. This layered approach aims to mitigate risks while also providing users with necessary information.
To address the urgent need for safety, Anthropic has implemented conservatively tuned safeguards. The company cautions that these measures may occasionally flag innocuous requests, but reassures that they trigger in less than 5% of sessions on average. Anthropic is actively pursuing the development of new models to fine-tune these safeguards further and reduce instances of false positives.
Monetarily, both Fable 5 and Mythos 5 come with a price point of $10 per million input tokens and $50 per million output tokens, which is a significant reduction compared to the previous Claude Mythos Preview, priced at $25 per million input tokens and $125 per million output tokens. This pricing strategy positions Anthropic as a competitive player in the rapidly evolving AI landscape.
Following the announcement, Microsoft revealed that Fable 5 has been integrated into Microsoft Foundry, allowing it to enhance functionalities within GitHub Copilot and the Foundry Agent Service. This collaboration underscores a significant partnership that aims to leverage Fable 5’s capabilities in real-world applications.
However, the introduction of these advanced AI models has not been without its critics in the cybersecurity sector. Some experts have expressed concerns about the implications of deploying models like Mythos 5 to the public. Sachin Puri, CEO of Network Solutions, recognized that the Mythos-class models signify a substantial advancement in the ability of AI to detect and respond to sophisticated cyber threats. Still, he emphasized the importance of integrating governance and accountability as these systems evolve.
Similarly, Jamie Moles, a senior technical manager at ExtraHop, raised eyebrows over the decision to launch a “Mythos-class” tool, referencing extensive warnings from safety experts regarding the potential risks to corporate security. He stressed the necessity for responsible deployment of such powerful AI technologies, ensuring they do not jeopardize the stability of essential systems that communities rely on.
Andrew Rubin, founder and CEO of Illumio, took a more provocative stance, asserting that the introduction of guardrails reflects an inherent lack of trust in the capabilities of the system itself. He articulated that constraints at the user interface do not address the fundamental mathematical principles behind the technology, cautioning that attackers are likely to exploit the underlying functionalities rather than the superficial constraints placed around them.
Moreover, Douglas McKee, director of vulnerability intelligence at Rapid7, contemplated the functional relationship between Fable and Mythos. He suggested that while Fable may serve as a versatile tool, Mythos would remain essential for security teams tasked with managing rigorous defense protocols. He urged Anthropic to ensure that Mythos-class access is restricted to vetted defensive teams to maintain pace with the evolving threat landscape.
As Anthropic pushes the boundaries of AI technology, the ongoing dialogue between innovation and safety will remain crucial.