Offensive AI agents have become capable enough that a company can attack itself about as well as an outsider can, according to a security founder who thinks that shifts the advantage to defenders. The claim comes from a builder of Strix, a product that runs such agents against a firm’s own applications and code. The founder’s view is that getting companies to use them, not the technology, is the limiting factor.
Why attacks are speeding up
The alarm is warranted. Coding agents let teams release code that no human has read, and attackers can hand much more of their routine work to automation. Microsoft recently described an intrusion that moved from a GitHub search to pulling Azure Key Vault data within five minutes. Cloudflare says daily requests from the AI pentesting and exploitation tools it tracks are up 30 times since October 1.
Two newer reports add to the picture. The FBI’s November warning covers a framework that lets low-skilled attackers pull off elaborate operations. Google’s latest report describes one that diffs code against public patches, then writes and launches exploits on its own. Its operator needn’t understand each step the agent takes.
The old imbalance
The argument rests on two long-standing gaps:
- A defender has to cover every surface, but an attacker needs only a single opening.
- The best attackers were rare, so few companies could afford to imitate them.
The author believes AI now equals top human hackers in nearly every case, and that attackers won’t hold much stronger tools than the ones defenders can buy. If that holds, a company can for the first time mimic its attackers closely, and the skill gap behind the old imbalance disappears.
What the defender starts with
A company owns its code and knows the intended behavior of its product. It can give an agent test logins for each user role and let it examine a change before release. An outsider must infer much of this, usually once the code is already live.
That leads to the central bet. A firm that runs current offensive AI agents on itself continuously, with a budget matching its attacker’s, ought to spot serious flaws first, since it has equal hacking power, richer information and a head start.
Why the numbers matter
The two data points show how fast the offense is moving. A five-minute path from a code search to cloud secrets leaves almost no time for a human response. A 30-fold jump in automated tool traffic means machine probing is now ordinary background activity. Both back up the author’s view that waiting for attackers to arrive is no longer a plan.
Where Strix comes in
Strix is meant for exactly this setup. After an agent discovers a flaw, it demonstrates the exploit and can draft a patch for engineers to check. The pitch is that any company can now field the capability once held by a handful of elite attackers.
The catch: getting teams to use it
This is the view of a vendor selling the tool. The support is a mix of incident reports and the author’s own reading of the field, and the claim that AI matches elite hackers is an opinion, not a measured result.
The practical point still stands. Feeding agents source code and pre-release changes only pays off if teams do it daily, and many don’t. The author places the remaining risk in that gap.
What security teams can do now
None of this removes the need for human judgment. Engineers still review each proposed fix before it ships, and someone has to decide which findings matter most when the agents report many at once.
Teams considering offensive AI agents can begin by giving them what an outsider lacks: repository access, test accounts per role and unreleased changes. Steady, repeated runs matter more than one test, because the incidents above unfold in minutes and the tooling traffic keeps climbing.
Defenders who delay may meet the same agents on the other side with less preparation.
Source: alexschapiro.com