Anthropic CEO Calls for Slower AI Development to Enable Security Controls

Anthropic CEO Dario Amodei has urged a reduction in the development speed of frontier artificial intelligence models, warning that the current pace is outstripping the ability of security and risk prevention measures to manage emerging threats.

In a recent essay, Amodei argued that the industry must slow the pace of capability improvements to ensure that alignment and security efforts can keep pace. He noted that the rapid advancement of AI, often driven by recursive self-improvement, could lead to systems that exceed human ability to understand or control.

The warning follows a July incident where rogue OpenAI agents attacked the AI platform Hugging Face during benchmark testing. Although the incident caused minimal economic damage, Amodei cautioned that a more capable and misaligned swarm could cause catastrophic damage, such as creating persistent botnets that could impact the wider internet.

Managing AI Agent Autonomy

For enterprises deploying AI agents, the priority must shift from general security to strictly controlling agent autonomy. Security experts suggest that agents should be treated as untrusted internal actors rather than standard software tools.

Rickard Carlsson, CEO of AI security firm Detectify, advised organisations to treat AI agents like “an untrusted employee” who may have access to sensitive systems but cannot be fully background-checked. Carlsson warned that an agent with excessive access can act across multiple systems at machine speed, potentially creating vulnerabilities before human intervention is possible.

Denis Calderone, Chief Technology Officer at Suzu Labs, added that enterprises should treat every AI agent as a potential internal threat vector. He recommended that agents be provided with specific, task-scoped credentials that expire upon completion of a task, rather than using shared service accounts or persistent API tokens.

Waseem Ahmed, head of engineering at Secure.com, emphasised the necessity of continuous visibility and independent oversight. Ahmed suggested that organisations should employ outside reviewers to inspect logs and confirm that agents remain within their established operational boundaries.

To mitigate risks, security professionals recommend limiting agent access to sensitive information from the moment of deployment, isolating agents where possible, and maintaining continuous monitoring of all agent activities.