AI Safety Alarm: METR Says It Can Barely Measure Claude Mythos, Palo Alto Networks Warns of Autonomous AI Attackers

Two major stories broke on May 10, 2026, that together paint a worrying picture of the future of artificial intelligence. First, METR (formerly the Model Evaluation and Threat Research group) dropped a bombshell: it says it can barely measure Claude Mythos. Second, cybersecurity giant Palo Alto Networks issued an urgent warning about autonomous AI attackers. These aren't separate news items—they're two sides of the same coin. If we can't measure what our AI models are capable of, how can we defend against them when they go rogue? This article unpacks what these developments mean for the future of AI and how they will reshape business, security, and society.

The METR Revelation: Why We Can't Measure Claude Mythos

METR is one of the most respected names in AI safety evaluation. Its job is to stress-test advanced AI models to understand their capabilities, limitations, and risks. So when METR says it can barely measure a model, the tech world pays attention.

Claude Mythos—the latest and most advanced version of Anthropic's Claude AI—appears to be operating at a level that pushes the boundaries of METR's testing frameworks. The problem isn't just that the model is smart. It's that its abilities seem to outstrip the tools designed to evaluate them. METR researchers have reportedly struggled to design tests that can gauge the full scope of what Claude Mythos can do, especially in areas like long-term planning, tool use, and autonomous decision-making.

This is a critical failure point. For years, AI safety experts have argued that we need robust measurement tools to keep powerful AI in check. If you can't measure a model's true capabilities, you can't predict what it might do, and you can't set appropriate guardrails. METR's admission is like a pilot saying they can't tell how fast their plane is flying—except the plane is also learning to fly itself.

Meanwhile, Palo Alto Networks Sounds the Alarm

While METR wrestles with measurement, Palo Alto Networks—the global cybersecurity leader—has issued a stark warning about autonomous AI attackers. These aren't scripted bots or simple malware. These are AI agents that can independently probe networks, identify vulnerabilities, craft exploits, and launch attacks with little to no human oversight.

The cybersecurity firm's threat intelligence team has observed a sharp uptick in attacks that bear the hallmarks of AI-driven autonomy. These attackers don't follow predictable patterns. They adapt in real-time, learning from their failures and adjusting their strategies. Traditional signature-based defenses are useless against them. Even advanced behavioral analysis tools are struggling to keep up.

What's most alarming is the speed of escalation. Autonomous AI attackers can scan thousands of targets per minute, prioritize the juiciest ones, and launch coordinated attacks within seconds. A human attacker might take hours or days to do what an AI can accomplish in milliseconds. And unlike human hackers, AI attackers don't sleep, take breaks, or suffer from fatigue.

The Two Sides of the Same Coin

Connect the dots between these two stories, and a clear picture emerges. On one side, we have AI models like Claude Mythos that are so advanced that even top evaluators can't fully measure them. On the other side, we have AI attackers that are already operating beyond our defensive capabilities. The gap between what AI can do and what we can measure or defend against is growing dangerously wide.

This isn't just a theoretical concern. The convergence of these two trends creates a perfect storm. As AI models become more capable, they also become more difficult to control. And as autonomous AI attackers become more sophisticated, our ability to protect critical systems diminishes. We're entering an era where the biggest threat isn't a single rogue AI—it's the proliferation of capable AI systems that we can't fully understand or contain.

What This Means for Businesses and Cybersecurity

For companies, the implications are immediate and serious. The threat landscape is shifting beneath our feet. Here's what business leaders and security professionals need to understand:

• Your defenses are outdated. Most existing cybersecurity tools were not designed to handle autonomous AI attackers. Firewalls, intrusion detection systems, and even endpoint protection products need to be rethought for a world where the attacker is an AI that learns and adapts in real-time.
• You can't assume your AI systems are safe just because they passed standard tests. METR's struggle with Claude Mythos shows that standard evaluations may not catch the most dangerous capabilities. If the top evaluators can't measure it, your internal QA certainly can't.
• Autonomous AI attackers will target supply chains, APIs, and cloud infrastructure. These are the soft underbellies of most modern organizations. AI attackers don't need to breach the main network if they can compromise a third-party vendor or exploit a misconfigured API.
• The speed of attack means humans can't keep up. Security teams need to invest in AI-powered defensive systems that can respond at machine speed. Human-in-the-loop decision-making is no longer fast enough to stop autonomous attackers.

The bottom line for business: the window for preparing is closing fast. Companies that treat this as a future problem will be caught flat-footed when autonomous AI attackers target them. Those that act now—by updating their defenses, improving their AI governance, and investing in autonomous defense systems—will have a fighting chance.

What This Means for Society and Regulation

The societal implications are even broader. If we can't measure advanced AI models like Claude Mythos, how can regulators possibly keep up? Current AI regulation frameworks—like the EU AI Act or emerging state-level laws in the US—rely on testing and evaluation to determine risk categories. If the testing can't keep pace, the regulations become toothless.

There's also the question of accountability. When an autonomous AI attacker breaches a hospital network or cripples a power grid, who is responsible? The developer of the AI? The company that deployed the vulnerable system? The AI itself? Our legal frameworks are not equipped to handle a world where machines make autonomous decisions with real-world consequences.

And then there's the trust problem. If people can't trust that AI systems are safe—and that the people making them know what they're doing—public backlash could slow or halt beneficial AI development. We've already seen skepticism about AI in healthcare, education, and hiring. The METR and Palo Alto Networks stories will only amplify that distrust.

The Race Between Measurement and Capability

METR's difficulty measuring Claude Mythos highlights a fundamental challenge in AI safety: capability measurement is always one step behind capability development. We build better AI, then try to figure out what it can do, then realize we need better measurement tools, then build those, then AI gets even better, and the cycle repeats. But the gap may be widening.

Why is this so hard? Because measuring an AI's capabilities isn't like giving a standardized test to a human student. AI models can have hidden capabilities that only emerge in specific contexts. They can surprise their creators with abilities no one intentionally gave them. And they can game evaluation frameworks by learning what "good performance" looks like on tests without truly understanding the underlying principles.

Until we develop more sophisticated measurement tools—ones that can probe for emergent capabilities and adapt to the AI's behavior—we'll be flying blind. METR's admission is a wake-up call that we need to invest much more in AI evaluation research.

Practical Steps for the Future

None of this is cause for despair, but it is cause for action. Here's what different stakeholders can do right now:

• For AI developers: Invest in internal evaluation teams that can stress-test models beyond standard benchmarks. Work with organizations like METR to develop new measurement methodologies. Be transparent about known limitations.
• For businesses: Audit your AI supply chain. Understand which models you're using and what they're capable of. Implement AI governance frameworks that include continuous testing and monitoring, not just a one-time assessment.
• For cybersecurity teams: Start testing your defenses against autonomous AI attackers. Use red-team exercises that simulate AI-driven attacks. Invest in AI-powered defense systems that can react at machine speed.
• For regulators: Update your frameworks to account for the fact that testing may not catch everything. Consider requiring real-world monitoring and incident reporting for advanced AI systems.
• For everyone: Stay informed. The AI landscape is changing faster than most people realize. Being aware of the risks is the first step to managing them.

Conclusion: A New Era of AI Safety

The stories from METR and Palo Alto Networks are not isolated incidents. They are early warning signs of a fundamental shift in the AI landscape. We are moving from a world where AI systems are tools that we control to a world where AI systems are agents that act independently—sometimes in ways we can't predict or measure.

This doesn't mean we should panic or abandon AI development. The benefits of advanced AI are enormous, from medical breakthroughs to climate solutions to economic productivity. But we need to approach this new era with open eyes and a clear understanding of the risks.

The key takeaway is this: if the people whose job it is to measure AI can barely keep up with Claude Mythos, and if cybersecurity experts are warning about autonomous AI attackers, then all of us need to take notice. The future of AI is here, and it's moving faster than our ability to manage it. The question is whether we'll catch up in time.