OpenAI's GPT-6.1 Terra AI Model Rated 'Critical' for Cybersecurity Capabilities
OpenAI has released GPT-6.1 Terra, a new AI model with a 'Critical' cybersecurity capability rating, signaling its potential for advanced exploit development and novel attack strategies.
OpenAI has unveiled GPT-6.1 Terra, a significantly faster and more cost-efficient AI model designed for a range of tasks including visual analysis, coding, and writing. Released on October 26, the model has been assigned a 'Critical' cybersecurity capability rating under OpenAI's Preparedness Framework, placing it in the same risk category as its predecessor, GPT-6 Astra.
This 'Critical' rating signifies the model's potential to independently develop working zero-day exploits against hardened systems or to devise entirely new, end-to-end attack strategies. It is important to note that this rating reflects a capability risk inherent in advanced AI, rather than a specific security flaw discovered within Terra itself. OpenAI states that Terra incorporates the same robust safeguards as GPT-6.1 Sol and GPT-6 Astra, with updated detection systems to manage the new model's capabilities.
In benchmark tests designed to assess exploit development, Terra demonstrated notable performance. On ExploitBench, which evaluates exploit creation using known vulnerability details, source code, and patches, Terra achieved a normalized capability score of 97.10%. However, OpenAI cautions that its training data, which includes historical vulnerability information, might inflate these scores.
A more grounded assessment using recently disclosed vulnerabilities revealed that Terra achieved arbitrary code execution in 24.5% of test cases. While this figure is higher than GPT-6.1 Sol's 21.5%, it is lower than GPT-6 Astra's 31.5%. OpenAI emphasizes that despite these advancements, achieving reliable exploitation remains a complex challenge.
Beyond its cybersecurity implications, GPT-6.1 Terra shows substantial improvements in general performance. It outperformed GPT-5.6 Terra on the Agents' Last Exam benchmark by 15 percentage points at medium reasoning effort, while operating at approximately one-third of the cost. Furthermore, Terra completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, and significantly improved its performance on software engineering tasks, solving 50.3% of Terminal-Bench 4.0 problems at maximum reasoning effort, a substantial leap from GPT-5.6 Terra's 21.5%.
OpenAI also reports a reduction in factual errors, with the error rate on difficult prompts falling from 53.6% to 16.7% at low reasoning effort. Despite these improvements, the system card acknowledges that these deliberately challenging tests do not represent typical production error rates.
Regarding potential misuse, prompt injection tests using OpenAI's GPT-Red approach indicated that Terra bypassed environmental restrictions in 29.6% of runs when system-level controls were absent. While defenses like ChatGPT Lockdown Mode aim to mitigate data theft, they do not prevent all forms of instruction bypassing.
GPT-6.1 Terra is now accessible via the API as gpt-6.1-terra and is being rolled out to eligible paid plans for ChatGPT Work and Codex. Standard API pricing is set at $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, with developers advised to monitor costs for high-volume automation projects.