Every capability a defender builds to find bugs faster is also a capability an attacker can build. Mandiant’s latest disclosure makes that trade-off impossible to ignore.
What the tool actually does
The Agentic Vulnerability Discovery Harness chains together specialised large language model agents under deterministic, human-defined orchestration rather than letting a single model freewheel over a codebase. An Explorer agent maps what a codebase does and where its risk areas are, specialist agents dig into authentication, authorisation and routing, other agents trace how user input flows through the application to dangerous sinks, and a validation stage has multiple agents challenge each other’s conclusions before a human consultant reproduces the exploit to confirm it is real. That last step matters. It is the difference between a tool that generates noise and one that generates findings a security team can act on.
The number that should worry every CISO
In one incident response engagement involving source code stolen from a corporate victim, AVDH found more than 100 critical, human-verified vulnerabilities in two days. Over ten months of internal use across tens of millions of lines of code, the harness has run thousands of analysis pipelines and produced twelve assigned CVEs so far, including flaws in widely used open-source projects, with more in active disclosure. Google built this to help its own defenders and incident response teams move at machine speed. The uncomfortable implication is that nothing about the underlying technique is exclusive to Google, and a well-funded attacker who has exfiltrated your source code has every reason to point similar tooling at it.
Why this changes the maths on a breach
For years, stolen source code was treated as a slow-burn risk, useful to an attacker mainly for competitive intelligence or as leverage in an extortion negotiation. AVDH demonstrates that a modern attacker with access to your code can realistically go from theft to a prioritised list of exploitable bugs within days, not months. That compresses the window a security team has to rotate credentials, patch dependent systems and assess blast radius after any code repository breach, and it means secure code review can no longer be treated as a periodic audit.
What this means for European teams
Most European mid-market and enterprise engineering teams do not have a Mandiant-scale AI vulnerability harness sitting behind their CI/CD pipeline, and most cannot build one from scratch. What they can do is close the gap that makes tools like this so effective against a stolen codebase in the first place, tightened access control reviews, dependency and input-validation hygiene, and automated code scanning integrated into the pipeline rather than run occasionally.
If a source code exposure would leave your applications open to the kind of rapid, automated exploitation Mandiant just demonstrated, contact Excello Digital. We help European engineering teams build secure development practices that hold up against attackers who are no longer working by hand.