What happened
On August 18, 2026, Google's Threat Intelligence Group published a blog post disclosing the architecture of a tool Mandiant researchers Alex Tselevich and Michael Maturi call the Agentic Vulnerability Discovery Harness, or AVDH. It is not a single AI prompt scanning code for problems. It is a sequential pipeline of specialized agents, built on Google's Agent Development Kit, that Mandiant has run internally for 10 months. The headline result: during a real incident-response investigation into stolen corporate source-code repositories, AVDH found more than 100 confirmed critical vulnerabilities in two days. Across its broader use, Google says the tool has covered tens of millions of lines of code, run thousands of analysis pipelines, and produced tens of thousands of findings, leading to 12 assigned CVEs so far, with roughly a dozen more in active disclosure. In a separate client engagement, it found a remote-code-execution flaw that gave a red team its initial access into the target application. The pipeline works in stages. An Explorer agent first profiles the codebase and builds a threat model, which a human reviews before anything else proceeds. Entry-point agents then scan every in-scope file for places user input can enter the application. Other agents gather the surrounding code, checks, and sanitizers around each entry point. Access-control and data-flow agents generate hypotheses about specific issues: missing authorization, privilege escalation, SQL injection, cross-site scripting, command injection, path traversal. A separate set of agents then stress-tests each hypothesis at a higher setting before a synthesis agent marks it confirmed, disproven, or rejected. Only after all of that does a Mandiant consultant manually reproduce the exploit with working proof-of-concept code. Nothing counts as a finding until a person has done that.
Why it matters for business owners
Google's own framing of this is direct: "Adversarial misuse of AI has increased the risk of data theft and extortion events, because when proprietary source code is exposed, defenders must scramble to identify and patch vulnerabilities while attackers deploy machine-speed AI tools against them." The same underlying AI capability that lets a defender find 100 bugs in two days lets an attacker do the same thing to code they have stolen or otherwise gained access to. That changes the math on source-code exposure. If your business, a contractor, or a software vendor you rely on ever has source code leaked or accessed by an outside party, whether through a breach, a departing employee, or a careless AI coding tool with broad repository access, the realistic window before that code gets systematically picked apart for exploitable flaws is now measured in days, not the weeks or months it used to take a human team. That is true whether the analysis is done by the attacker or, in Google's case, by a defender racing to patch first.
What owners should not misunderstand
It is easy to read '100 critical vulnerabilities found by AI in two days' and assume any AI-branded security tool can now do this. That is not what Google disclosed. AVDH sits on top of roughly a decade of Mandiant's proprietary consulting playbooks and custom-built benchmark codebases, refined over 10 months of internal engineering by one of the world's largest security research teams. That is not a resource most companies, including most mid-size businesses, can replicate in-house, and Google is not describing an open capability. It is describing the research behind a product it is folding into its existing CodeMender scanning tool and its always-on AI Threat Defense platform, meaning most businesses will end up buying this capability from a vendor rather than building it. The more important detail is the human validation gate. Google states plainly: 'We encourage network defenders considering implementing similar vulnerability discovery harnesses to manually validate findings.' Even the team that built this tool does not trust its raw output on its own. A vendor pitching you an 'AI-powered' or 'AI-verified' security product may or may not have anything resembling that validation step. The word 'verified' in a sales conversation and the word 'verified' in Mandiant's actual process are not automatically the same thing.
The operational lesson
AI is compressing the timeline on both sides of this problem at once. Attackers can move from stolen code to a working exploit faster than before. Defenders, at least ones with Google's resources, can move from exposure to a patched fix faster too. For a business that is not building either capability in-house, the useful question is not 'do we have an AI security tool.' It is two narrower questions: how exposed is our source code and our vendors' source code right now, and when we buy a tool that claims to find or verify vulnerabilities with AI, what specifically happens between the AI's output and something we act on.
What a serious business should do next
Ask any current or prospective security vendor exactly what 'AI-verified' means in their product. Is there a mandatory human review step before a finding is reported or acted on, or does the AI's output go straight to you as-is? Get a specific answer, not a marketing line. Map who and what currently has access to your source code: employees, contractors, outside agencies, and any AI coding assistant or automation tool with repository access. The relevant question is no longer just whether that access is appropriate. It is how quickly your code could be turned into a working exploit if that access were misused or that code were exposed. Do not add an 'AI security' line item to your budget because the category feels urgent. If code security genuinely matters to your business, given your industry, your data, and what a breach would actually cost you, evaluate a specialized vendor on their validation process, not on how many findings their tool claims to produce. This is a case where using a specialized vendor makes more sense than trying to build something internally, but the selection criteria should be the human check behind the AI, not the size of the number in the demo. If you do not currently know how exposed your own source code is, or whether your existing tools have any human validation step at all, that gap is worth closing before you spend on anything new.
The Atlacis view
A vendor's AI-powered security claim is not something a business owner should have to take on faith. Atlacis helps owners cut through what an 'AI-verified' or 'AI-powered' claim actually means in a specific product, map where their real code and data exposure sits, and decide whether a security tool matches the risk they actually carry, rather than adding a subscription because the headline sounds urgent.
The short version
- On August 18, 2026, Google's Mandiant team disclosed AVDH, an AI agent pipeline it has run internally for 10 months, which found more than 100 confirmed critical vulnerabilities in stolen corporate source code in two days during a real incident-response case.
- Across broader use, the tool has covered tens of millions of lines of code and led to 12 assigned CVEs so far, with about a dozen more in active disclosure.
- Every finding still passes through a mandatory human validation step, a Mandiant consultant manually reproducing the exploit, before it counts. Google's own team does not trust the AI's raw output on its own.
- This capability rests on roughly a decade of proprietary Mandiant playbooks and 10 months of internal engineering, resources most businesses cannot replicate. Google is folding it into vendor products (CodeMender, AI Threat Defense), meaning most companies will buy this capability, not build it.
- The same class of AI capability that helps a defender find bugs fast can help an attacker exploit stolen or exposed source code fast, shrinking the safe window after any code exposure.
- Before buying any AI-branded security tool, ask exactly what 'AI-verified' means in that product and whether a human confirms a finding before it is acted on.
Where ATLACIS can help
Sources
- Google Cloud Blog (Google Threat Intelligence Group): Staying Ahead of Adversarial AI Through Agentic Source Code Review (Alex Tselevich, Michael Maturi, August 18, 2026)
- Help Net Security: Google's AI security agents found 100+ critical software vulnerabilities in just two days (Sinisa Markovic, August 19, 2026)
- SecurityBrief: Google Cloud unveils AI harness that finds flaws fast (Mark Tarre, August 18-19, 2026)