Anthropic AI controversy raises cyber trust concerns

Security researchers and vendors warn that recent controversies involving Anthropic's AI systems point to deeper questions about attribution, national security, and the rapid spread of offensive cyber tools.

The debate has intensified after reports that Claude included hidden markers aimed at Chinese users, along with a separate clash between Anthropic and the US government over access to its Mythos Lite Fable 5 model.

Specialists say the issues now go beyond a single provider or country, cutting across how AI models are built, monitored, and constrained.

Anthropic has come under scrutiny over claims that Claude contained spyware-like elements that identified requests from China or treated them differently. Reports alleged that prompt instructions inside the model carried distinctive "fingerprints" that could enable attribution of usage linked to Chinese infrastructure.

Gabrielle Hempel, Security Operations Strategist at Exabeam, said embedding identifying markers in software has a long history in cybersecurity and product design.

"This is less of a China issue and more of an AI attribution and trust issue. From a security perspective, I actually don't find the existence of these markers very surprising. For decades, we've been building canaries, watermarks, telemetry beacons, and attribution mechanisms into software. As AI models become strategic assets, it makes sense that vendors are looking for ways to determine who is using their systems, where requests are originating, and whether users are attempting to circumvent controls," said Gabrielle Hempel, Security Operations Strategist at Exabeam.

Critics argue that such embedded signals can operate without users' awareness, blurring the line between technical risk management, surveillance, and geopolitical monitoring.

Hempel said the trend raises broader concerns for the AI sector, including how much visibility users have into model behavior and what role governments may play in setting limits on tracking and controls.

"What concerns me more is the precedent this establishes. If frontier AI providers are embedding hidden attribution signals inside prompts and contexts, we're moving toward a world where AI systems themselves become active participants in supply chain intelligence and geopolitical competition. It fits the broader trend we've been watching, that AI capability development is increasingly inseparable from national security concerns. The conversation is no longer simply about who can build the best model, but who can protect, attribute, and control access to those models," Hempel said.

The tension between security, transparency, and access has also played out in the saga around Mythos Lite Fable 5. Anthropic restricted the model after White House concerns over national security and misuse risks, then restored access after a two-week standoff.

Mythos Lite Fable 5 drew attention in cybersecurity circles because of its reported ability to find software vulnerabilities and support exploit development. US officials warned that widely accessible tools with this level of performance could strengthen hostile states or criminal groups.

For some practitioners, the episode underscored the limits of restricting any single frontier model. Competing tools are emerging from research groups and vendors across multiple jurisdictions.

"Restricting one frontier model doesn't restrict the capability for long. We're already seeing comparable performance emerge elsewhere, from Japan's Sakana AI with Fugu, which orchestrates multiple frontier models, to China's open-weight GLM-5.2, which is reportedly performing at Mythos-level on vulnerability detection. The model race will keep shifting, but the underlying capability is only becoming more widespread. For security teams, there are really only two things that matter: AI can now find and exploit vulnerabilities at a scale and speed no human team can match, and that capability is increasingly available to anyone, including attackers. The challenge now is helping defenders move at the same speed as the models writing code and discovering security flaws. In this new era, organizations need to focus less on which AI 'wins the race' and more on leveraging and operationalizing AI to make security as agentic as AI-powered development and AI-powered attackers," said Ronen Shetelboim, CMO at Cycode.

Security leaders now face parallel concerns: how AI models may track and classify users, and how to adapt to an environment in which offensive-grade AI for vulnerability discovery is spreading across borders and product lines.