Anthropic on AI agents in cybersecurity: what they can do and where they fall short

Anthropic has published research on applying AI agents to cybersecurity tasks. Edgar Sipki, a Habr author and founder of easyp & sipki tech, set out to pick the document apart and ask an uncomfortable question: how far can these agents be trusted in practice?

What Anthropic says

The company tested Claude agents on a broad range of information security tasks: static code analysis, vulnerability discovery, threat modeling, and infrastructure security assessment. In typical scenarios, the agents performed above the level of an average specialist. They processed large codebases faster and spotted common vulnerability patterns that are easy to miss in a manual review done under deadline.

Areas where agents already deliver practical value:

• Static code analysis: finding SQL injections, XSS, insecure dependencies, and hardcoded secrets
• Automatic threat modeling for new services
• Faster penetration testing: the agent maps the attack surface while the specialist focuses on complex vectors
• Detailed risk reports and recommendations for prioritizing fixes
• Monitoring codebase changes for security regressions

On paper, this sounds convincing. A closer look complicates the picture.

Where agents fall short

The main problem is quality on edge cases. Agents hallucinate vulnerabilities that do not exist, and at the same time miss real problems hidden behind unusual code or specific business logic.

In cybersecurity this matters especially. A false positive wastes the team's resources on investigating a nonexistent threat. A false negative leaves a real hole open and creates a false sense of security. The second scenario is worse than no check at all: it lulls teams into complacency.

Another weakness is limited system context. The agent sees only what it has been given. Vulnerabilities that depend on the interaction of several components, on deployment specifics, or on a particular cloud environment often escape it, because they require understanding the whole architecture, not individual files.

Access rights are a separate problem

Here is a serious question that rarely appears in marketing materials: what permissions does an agent need to work effectively in a security context? Full infrastructure scanning requires elevated privileges. And elevated privileges are themselves an attack vector: if the agent is compromised or makes an error with consequences, the scale of the problem grows quickly.

Anthropic recommends the principle of least privilege and isolated environments for agentic security tasks. But setting this up properly takes additional engineering work, and under tight deadlines it is often skipped. That creates exactly the hole the agent is meant to close.

"Agents are not a replacement for security specialists but a tool for speeding up their work" is the key thesis of the

Anthropic research.

What this means

AI agents already help with routine security tasks: standard code review, initial scanning, and report preparation. But giving them autonomous authority over critical infrastructure is still premature.

The main conclusion, which Edgar Sipki shares: agents change a security team's toolkit, not its composition. A human in the loop remains mandatory, especially where the cost of an error is high.