Mantic forecasts AI hacking costs falling to about $1 per task by 2027

Mantic’s new cybersecurity forecasts suggest open-weight AI models could reach today’s frontier hacking benchmark by late 2027, while the estimated cost of Mythos-level hacking could fall to about $1 per task. The projections turn a noisy argument about AI and cyber risk into specific claims readers can watch, though they remain forecasts, not observed outcomes.

Mantic Watch analysis

What happened

In a post published on 27 October, Mantic describes forecasts made with its AI forecasting system. It estimates that the best open-weight model could match today’s reported frontier score on ExploitGym, a fixed 869-task test, around the end of September 2027. Mantic also forecasts the estimated cost of reaching its chosen “Mythos-level” benchmark at about $1 per task by the end of 2027, down from its estimate of $3.01 for a qualifying open-weight model today.

The post gives other forecasts too. Mantic assigns about a 31% chance to the Artificial Analysis Cyber Index reaching a score of 80 by 30 September 2027. In a scenario where defensive AI becomes very strong, it forecasts a higher median annual count of exploited zero-days than in its baseline: 227 rather than 190.

Why it matters

These estimates make the disagreement over AI-enabled cyber risk more testable. If capable open-weight models become cheaper and catch up with frontier systems, more people may be able to use them for security work, and potentially for attacks. But benchmark scores and estimated task costs do not tell us how often real vulnerabilities will be exploited or how much damage will follow.

Mantic’s figures also resist the tidy assumption that stronger defensive AI automatically means fewer attacks. Its scenario forecasts more recorded zero-days, while noting that discovery, disclosure and tracking affect the count. The forecasts are conditional estimates, not evidence that this outcome is already happening.

Our read

The value here is less “AI will end cybersecurity” than a set of numbers that can be revisited as models, costs and incident data change. Mantic has a commercial interest in forecasting, so readers should treat its projections as the company’s analysis, not a neutral verdict. Still, putting assumptions and measurable targets on the table beats another round of cyber-apocalypse fog.

What to watch

• Whether open-weight models approach the cited ExploitGym score by late 2027.
• Whether estimated costs move towards Mantic’s $1-per-task forecast.
• How the Cyber Index and real-world vulnerability and incident data change.
• Whether later Mantic forecasts revise these estimates as new evidence arrives.

Discussion spark: If stronger AI helps defenders find flaws but also makes capable hacking cheaper, which outcome should governments and security teams prioritise measuring: vulnerabilities fixed before exploitation, or attacks that succeed?

not affiliated with or endorsed by Mantic