Dario Amodei, CEO of Anthropic, testifies during a US Senate Privacy, Technology, and the Law Subcommittee hearing on "Oversight of A.I.: Principles for Regulation," on Capitol Hill in Washington, DC, on July 25, 2023. (Photo by SAUL LOEB / AFP) (Photo by SAUL LOEB/AFP via Getty Images)
AFP via Getty Images
Anthropic’s Claude Mythos Preview has sparked concern in the U.S. and globally over AI safety. Debate has spread from Wall Street and Washington, D.C., to financial institutions in Europe. Anthropic is withholding it from public release, citing the model’s apparent ability to autonomously exploit previously unknown cybersecurity vulnerabilities.
AI is already a kind of Pandora’s box. Its impact can scale at extraordinary speed because its outputs are automated, reproducible, and easily multiplied. That does not make AI the same as a nuclear weapon. But it does make it a system-level risk. Once highly capable models are widely accessible, misuse can spread fast across industries and institutions.
But commercial pressure may be moving faster than governance. The erosion of safety capacity at major AI companies has drawn scrutiny. OpenAI’s reported shutdown of its Mission Alignment team earlier this year and the disbanding of dedicated AI safety team in 2024 were almost like racing a horse without a bridle. When safety functions shrink as model capability grows, the technology becomes more vulnerable to malicious use. Public anxiety is only natural.
To reduce cybersecurity risks, Anthropic launched Project Glasswing. It is a coordinated vulnerability disclosure effort involving Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The goal is to let major infrastructure providers use the model’s predictive power to find weaknesses in cloud systems and patch critical software bugs. The idea is to fix those problems before the model, or a similar adversarial system, reaches a broader and less regulated public. This marks a shift in industry priorities. The race in AI is no longer only about capability. It is also about who can secure the systems that capability may threaten.
As AI capability rises, so do the risks across industries. In finance, models with Mythos-level reasoning could help simulate or execute complex market manipulation, evade fraud detection, or automate the discovery of institutional weaknesses. In manufacturing and commerce, advanced AI could identify and exploit supply-chain bottlenecks, causing delays, disruptions, or theft at scale. In universities and research institutions, the threat extends to proprietary research data, internal networks, and AI-assisted social engineering attacks against administrators and faculty.
As AI becomes a general-purpose tool for productivity, it also becomes a general-purpose tool for sabotage. The same flexibility that makes it commercially valuable also makes it highly adaptable for malicious use.
The 2026 Stanford AI Index Report, released this month, highlights a sharp increase in AI adoption in medicine. It notes a significant rise in AI uses for clinical documentation, medical imaging, and diagnostic reasoning. That growth may improve efficiency. But it also expands the attack surface for public health if mis-deployed.
If a model like Mythos were used to corrupt medical databases, manipulate diagnostic systems, or generate inaccurate pharmacological guidance, the damage would go far beyond an ordinary data breach. It would create a direct threat to patient safety. As healthcare systems become more dependent on AI-mediated workflows, the prospect of adversarial medicine becomes harder to dismiss. In such cases, bad actors could manipulate AI outputs to cause harm, create confusion, or extort hospitals. In that context, identity verification and access controls look less like optional friction and more like core infrastructure.
The possibility that high-capability models could enable such harms has accelerated a shift toward mandatory identity verification. Anthropic now requires government-issued identification and biometric live selfies from users seeking access to certain high-risk functions. The company frames this as a matter of platform integrity, arguing that responsible use of powerful technology begins with knowing who is using it. To address privacy concerns, Anthropic says the verification data is not used to train models and is not shared with third parties for marketing or advertising.
Physical ID checks for AI use may feel like a major change in user experience. But they also extend a much older industry practice. For years, technology companies have relied on passive forms of verification. Google sign-ins for Gemini, internet service registrations, and the metadata tied to email accounts, devices, and digital purchases already provide dense identity signals. Those systems have long supported public-safety functions and commercial monetization. Explicit identity checks for advanced AI models formalize that trajectory. They move the industry from background identification to foregrounded, bank-grade authentication.
Anthropic’s Project Glasswing brings together major cloud providers and cybersecurity companies. But it does not yet appear to meaningfully include public institutions or policymakers.
The gap matters. Innovation is vital to economic competitiveness. But safety remains a precondition for lasting growth. Governments once built legal frameworks for cybersecurity and data privacy. They now need to update AI safety laws and regulatory mechanisms to match the capabilities of new models. Without that public framework, too much of the burden will fall on private firms whose incentives do not always align with the public interest. The real question is whether institutions can move fast enough to govern it before the technology outruns existing controls.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。