IBM is putting frontier AI models on the red team. The company on Monday launched X-Force Advanced Proactive Security, a new service that lets its human security testers direct frontier AI models — via OpenAI's Daybreak program and other best-of-breed solutions — to probe enterprise networks, applications and infrastructure for weaknesses before attackers do.
The idea is to give security leaders something they rarely have: continuous, validated evidence of what an attacker could actually do inside their environment. Traditional penetration testing still runs on a snapshot model — scoped, scheduled and paid for once or twice a year. "Between tests, most of the estate goes unchecked, and attackers already using AI are ready to exploit that gap," wrote Ray Sims of IBM's X-Force Offering Strategy in the launch announcement.
Humans in charge, AI on the tools
IBM is careful to frame this as expert-directed testing, not automation for its own sake. Engagements start from business risk, not from the model: X-Force assesses what each application holds and can reach, how exposed and critical it is, and what regulators expect — then recommends how deep the testing should go, how often it should run and how much human oversight it needs. The client makes the final call, and IBM updates its recommendation as the environment changes.
In expert-led tests, frontier models assist with reconnaissance, attack-surface mapping, exploit development, ranking results and drafting reports. The human tester chooses which model handles each task based on measured performance and validates AI output where the stakes demand it. IBM says its testers can find business-logic flaws that automation typically misses.
Autonomous tests on a schedule
For lower-risk work, the machines go solo. External network penetration tests and low-risk web application tests can run autonomously on a schedule the client sets, with the option to add human validation or request retests on demand. X-Force testers set the policy and review anything escalated before it reaches the client — the aim being, in IBM's words, "an ongoing answer instead of an annual snapshot" of external exposure and application risk.
Everything runs on one platform: a single portal to set up and request tests, one deduplicated view of findings per application and one way to retest each fix. Findings flow into the ticketing, vulnerability and exposure-management tools teams already use. Commercially, clients get one testing budget covering the whole X-Force offensive catalog, with each test in the published catalog carrying a known cost — so adding depth doesn't require a new contract or scoping exercise.
Fighting AI with AI
The launch sits inside IBM's broader autonomous-security push, alongside the Autonomous Threat Operations Machine (ATOM) agents already used for SOC analysis, threat intelligence and threat hunting, plus the recently announced AI Cyber Resilience program. It is also a sign of how fast the industry is moving: IBM isn't just defending with AI — it is hiring AI to attack first.
As AI-driven attack tools get cheaper and faster, the old model of an annual pentest is starting to look like a security camera that only records on Tuesdays. IBM's bet is that the answer isn't more tools generating more unvalidated findings — it's controlled, continuous testing with humans still holding the steering wheel.
By Hamza Faisal
