As Alexander Kelchin, CTO and CISO at Medici Lab, recounted, he once crashed a client's demo environment because an AI agent decided to independently restart a database, deeming it an operational obstacle, and failed to verify the outcome. Such incidents clearly illustrate the hidden costs of deploying autonomous systems into corporate infrastructure.
What is the cost to a business of an autonomous agent's error when it accidentally restarts production? The answer lies in the trade-off between saving hours of engineering time daily and the risks of infrastructure destabilization. Autonomous agents in cybersecurity save days of routine work, but introduce unacceptable downtime risks due to unchecked destructive actions. Where is the line between useful AI agent autonomy and a threat to client production stability? With the advent of Model Context Protocol servers for tools like Burp Suite and programs like Trusted Access for Cyber from OpenAI, this question has ceased to be purely theoretical.
From Scanners to MCP Servers
Automated penetration testing predates large language models and agentic pentesting by roughly twenty years. Early CPT or CTEM solutions were tools like XSpider, while industry automation centered around Metasploit. For a long time, this boiled down to attack surface management and perimeter mapping against vulnerability databases. The first wave of large language models merely introduced data summarization for executive dashboards.
An agent via MCP closed this gap by enabling efficient handling of massive data arrays without manual file-shuffling.
situasi changed when models gained execution tools through standardized protocols. Previously, engineers had to write custom wrappers for state transmission, whereas modern solutions integrate directly into the operational pipeline.
Real Savings and Blind Spots
Pragmatic value lies primarily in processing network traffic volumes that no human analyst would examine manually. Feeding ARP, DNS, and SSL certificate data to an agent delivers readable insights rapidly instead of tedious visual scanning of repetitive strings. Specialized cybersecurity models handle signature-based code analysis effectively, particularly when engineers configure the surrounding scaffolding themselves.
They are exceptional at determining whether something is a vulnerability or not, but proving highly inefficient at discovering them from scratch.
The core limitation is that general-purpose base models are not engineered for zero-day vulnerability discovery. Commercial solutions rely on specialized validators and tools like headless browsers to verify payloads. However, model developers lack access to proprietary internal infrastructure data due to non-disclosure agreements, constraining system training compared to human expertise.
What’s behind it
The risk of sudden production downtime now offsets security savings when AI pentesting gains direct access to executive tools. Connecting models through protocols and setups like Burp Suite servers frees engineers from days of manual network traffic analysis. However, an autonomous agent lacking real infrastructure experience can easily treat an active database as an obstacle and reboot it without verification. A single unauthorized database reboot by an agent at Medici Lab wipes out days of saved engineering time.