Somewhere right now, a security researcher is watching a large language model do in minutes what used to take a skilled human tester days. The model reads through a codebase, spots a suspicious pattern in how input gets handled, and proposes a working exploit. It is not always right. But it is getting better, and that fact has quietly become one of the most consequential storylines in cybersecurity.
The idea that AI could assist with offensive security work is not new. What has changed is the pace. Recent industry commentary, including a widely shared piece from Cybersecurity Insiders, frames the question bluntly: as models grow more capable, can defenders keep up? The answer is complicated, and it depends heavily on which side of the arms race you are measuring.
On the attacker’s side, the advantages are real but often overstated. Modern models are genuinely useful at reconnaissance, at drafting convincing phishing lures, and at scanning open-source code for known vulnerability patterns. They lower the skill floor, which means people with limited technical depth can attempt things they previously could not. That is a meaningful shift. But models also hallucinate vulnerabilities that do not exist, struggle with complex multi-step exploitation, and rarely produce a clean, reliable attack chain without human guidance. The realistic threat today is not a fully autonomous hacker. It is an accelerant for people who already know what they are doing.
Defenders, meanwhile, are not standing still. Security vendors have spent the past two years embedding machine learning into everything from endpoint detection to log analysis, and the newer generation of models is being pointed at the same problems. The pitch is straightforward: if AI can find flaws faster, then AI should be able to triage alerts, prioritize patches, and spot anomalies across enormous volumes of telemetry that no human team could review manually. For overstretched security operations centers, that is less a luxury than a survival strategy.
There is a catch, though, and it is the same one that shows up across enterprise AI deployments. Models are only as good as the context they are given, and security work is deeply context-dependent. A tool that flags suspicious behavior without understanding your environment will drown analysts in false positives. The most useful deployments tend to be narrow: a model trained on a specific class of vulnerability, or one that assists with threat intelligence summarization rather than making autonomous decisions. Broad promises of “AI-powered security” tend to disappoint.
Then there is the uncomfortable question of the models themselves. The same systems being deployed to defend networks are also potential attack surfaces. Prompt injection, data poisoning, and model exfiltration are no longer theoretical concerns for security teams. A defender who automates patching decisions through an LLM has introduced a new dependency that an adversary might try to manipulate. Researchers have been warning about this for a while, and the warnings are starting to show up in real incident reports rather than academic papers.
Regulators and standards bodies are paying attention, though slowly. Frameworks for evaluating AI systems in high-stakes domains are emerging, but security tooling moves faster than the guidance meant to govern it. That gap leaves enterprises making judgment calls on their own, often without clear benchmarks for what “AI-assisted security” should actually deliver. Buyers would be wise to treat vendor claims with the same skepticism they apply to any other marketing.
The most grounded view may be that this is less a revolution than an acceleration of an old dynamic. Attackers and defenders have always borrowed each other’s tools. Fuzzers, once exotic, became standard on both sides. Automated scanning did the same. AI is following that pattern, just faster and with more hype attached. The organizations that fare best will likely be the ones that treat these tools as force multipliers for skilled people rather than replacements for them, and that invest in the unglamorous basics like patching, logging, and access control.
None of that makes for a tidy headline. But it is probably closer to the truth than either the doom scenarios or the vendor promises. The race is real. It is just being run by humans who happen to have better tools now, and the outcome will depend less on any single model than on how carefully those humans choose to use it.