Analysis · 2026-10-15

In cybersecurity, whose side is AI on: attack or defense?

Security companies say AI agents have greatly reduced the time they spend on analysis, investigation and patching. Researchers and the heads of AI labs warn that the same capabilities could be used for attacks. This article puts this week's case studies and warnings side by side, looks at which way the published evidence leans, and lists what practitioners should check.

Key points

In case studies published by OpenAI and AWS, companies said AI cut the time for malware analysis, threat investigation and patch suggestions from weeks to hours, and from hours to minutes.
Tamnoon hands everything from detection to remediation to agents, but added a separate stage for investigation and safety assessment before anything touches the live environment.
According to a news report, Microsoft credited a researcher who used Zhipu's GLM for 23 vulnerability reports at once.
More than 400 researchers, including Hinton and Bengio, along with Demis Hassabis, called for an international oversight body and standards for AI with cyber capabilities.
Most of the published figures on defensive gains come from the vendors themselves, and the measurement methods were not disclosed.

On the defensive side, what got faster, and by how much?

Most of this week's security news was about defensive automation. The scale of the time savings stands out: weeks become hours, and work that took hours is done in minutes.

According to an OpenAI announcement source, SentinelOne used OpenAI Daybreak to cut malware analysis that took weeks down to hours, which freed up capacity to investigate more threats. However, the published material does not describe the measurement method or scope. In the Block case study OpenAI presented source, OpenAI also said the security team investigates threats faster and suggests patches within minutes. The AWS blog source featured a case study from Orca Security. Orca Security built Discovery AI, an agent that turns plain-English questions into Sonar queries, and said it reduced query costs by 90%, input tokens by 93% and latency by 75% compared with its initial baseline.

All three cases have one thing in common: AI shortened investigation and query steps that people used to do by hand. However, all the figures come from the companies selling the tools. For practitioners, knowing which step got shorter is more useful than the 'weeks to hours' figure itself.

What changes when agents handle remediation as well as detection?

Detection findings already exceed what people can handle. The AWS blog source points out that CNAPPs (cloud-native application protection platforms) and scanners produce findings by the thousands. According to the post, most findings go unaddressed not because they are hard to fix, but because it takes time to investigate the context, find the owner, and confirm that a fix will not break the production environment.

Tami, which Tamnoon built on Amazon Bedrock, targets this gap source. Tami investigates findings on its own, assesses whether remediation is safe, and then carries out fixes with guardrails (safety mechanisms) and governance in place. The work is split across four specialized agents that handle tasks such as investigation and safety assessment. The investigation agent builds its plan in a read-only investigation stage. Tamnoon said it reduced critical cloud exposures by more than 90% over 90 days, with zero production incidents.

The key point of this case is arguably less about model performance than about how the stage 'before anything touches the live environment' was designed. To automate remediation, you first need to decide how much permission to give the agent and how to roll back when something goes wrong.

How do the same capabilities look from the attack side?

The ability to find vulnerabilities means different things depending on who uses it. For someone reporting vulnerabilities, it is a defensive tool; for an attacker, it is a weapon. This week brought news that illustrates this dual use.

According to the headline of an InfoQ China article source, Microsoft credited someone for reporting 23 vulnerabilities at once. Behind this were reportedly Zhipu's GLM model and a top-tier white-hat hacker. The collected material does not say which products the vulnerabilities were in or how GLM was used. There was also a report from The Information source, cited by AI Times. It said OpenAI is in early talks to acquire Trail of Bits, a company specializing in analyzing and fixing vulnerabilities in open-source software. Trail of Bits was founded in 2012 and has more than 40 employees. The deal size was not disclosed, and the talks may not result in a deal.

The GLM case can be read as a sign that vulnerabilities found with LLMs are useful enough to earn official credit from a major vendor. Seen the other way, it also means attackers can find vulnerabilities at the same level.

Why did calls for international oversight grow in the same week?

While defensive results were being published, risk warnings came out as well. According to an AI Times report source, more than 400 researchers from 34 countries, including Geoffrey Hinton and Yoshua Bengio, issued an open letter. The letter called for an international body to oversee containment standards and incident reporting for models with advanced cyberattack capabilities. The signatories argued that voluntary commitments from companies alone are not enough. According to an Electronic Times report source, Demis Hassabis also warned that if safety measures for AI agents are inadequate, serious security incidents could occur within two years, and called for international standards.

Opinions differ on who bears the risk. According to the headline of an article shared on Reddit source, an Anthropic executive argued that open-source models pose greater cybersecurity risks. However, who said it and on what basis could not be confirmed. Taken together with the report that the GLM family, which publishes its weights, was used for vulnerability reports source, the debate over the risks and usefulness of open models is arguably still unresolved.

So this week's evidence alone does not settle whether attack or defense has the advantage. The defensive side has public case studies and figures; the attack side has warnings and claims. That said, this asymmetry may come from a structure in which defensive results are easy to promote while offensive results stay hidden.

What practitioners can try

Break down the time spent working through the security backlog into investigation, owner identification, remediation and verification, measure each, and first find the stages AI agents can actually shorten.
When testing remediation automation, separate the read-only investigation stage from the execution stage, as Tamnoon did, and design agent permissions around the principle of least privilege and rollback procedures.
For vendor figures such as 'weeks to hours,' ask for the measurement method and scope, then re-verify them in your own environment using the same metrics.

Sources

An analysis written by AI from our briefing stories and official sources. The interpretation is one reading of the evidence; please check the originals for accuracy.