Frontier AI is making cyberattacks faster and cheaper, Ensign says

Frontier AI is no longer just a defensive headache. According to Ensign InfoSecurity, it is starting to change the basic economics of cybercrime.

In a 24 August 2026 newsroom post, the Asia Pacific cybersecurity company said its latest Cyber Threat Landscape Report found frontier AI was lowering the barriers to sophisticated cyberattacks as ransomware, data theft and hacktivism continued to rise across the region.

That is the primary claim from Ensign. Secondary accounts from Cybersecurity Asia and CRN Asia, both published in August 2026, add detail on the testing behind it.

Cybersecurity Asia reported that Ensign assessed 10 generally available frontier AI models using its proprietary AI Cyber Range Assessment. The models were asked to act as advanced cyber threat actors and were presented with eight attacker objectives.

The report, as described by Cybersecurity Asia, found the models could execute multiple stages of a realistic enterprise cyberattack chain. Ensign’s assessment said OpenAI’s GPT-5.6 Sol, Anthropic’s Claude Opus 4.8 and Z.AI’s GLM-5.2 were the strongest performers, achieving high success in seven of the eight objectives.

The more uncomfortable bit for defenders is cost.

CRN Asia reported that Ensign found Z.AI’s GLM-5.2 matched the offensive performance of GPT-5.6 Sol and Claude Opus 4.8 while costing roughly one-fifth as much to operate. Cybersecurity Asia also reported that GLM-5.2 delivered comparable offensive performance to GPT-5.6 at about one-fifth of the operating cost.

CRN Asia said Ensign had been evaluating models from both Eastern and Western developers since April 2026. The test group was reportedly drawn from more than 150 candidates, with the 10 models scoring above the 95th percentile on benchmarks such as CyberGym and ExploitGym taken into deeper testing.

The test environment, according to CRN Asia, was built to resemble a university network, with an online learning portal at the front and staff and student login systems behind it. Each model was given the same eight objectives and ran 16 times, with the network reset before each run.

Not everything worked for the AI attackers.

Cybersecurity Asia reported that initial access was shown to be trivial across the models, while at least half struggled to steal credentials and move between systems consistently. It also said none of the models could confidently evade detection tools.

CRN Asia reported that breaching the portal worked in every run across every model, while bypassing endpoint detection and response at runtime was the weakest result, with no model reaching high success.

Jeremy Moke, EnSOC director at Ensign InfoSecurity Malaysia, told CRN Asia: “In our testing, comparable offensive capability was available at roughly a fifth of the price, and defenders should assume the pool of adversaries will keep growing.”

For security teams, the message is blunt. Ensign’s findings suggest AI is not replacing existing attack methods. It is making them cheaper, faster and easier to scale.