COMMENTARY: After OpenAI published its assessment of Astra, it’s latest GPT model, the tool was categorized as the first model from OpenAI that meets the company’s critical cybersecurity threshold.Unfortunately, this also means that a well-equipped adversary could find previously unknown security flaws and autonomously exploit them across various well-protected organizational systems.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]
The company’s decision to restrict the most advanced cybersecurity capabilities under Astra only to those with access through OpenAI's Daybreak Blue should serve as a reminder of how powerful frontier AI models have become.Several autonomous, unsanctioned AI-driven attacks have played out in the public eye over the past several months, leading the industry to proceed with caution when releasing models like Astra. When a company determines that models are too capable to release to the public, security teams get a preview of the offensive power that will eventually reach adversaries with no restrictions.Race against the clockAs frontier AI models are shrinking the time between vulnerability discovery and exploitation to near-zero, security teams have struggled to keep pace. The challenge was evident in our recent report, where 70% of security leaders said an undetected threat has already resulted in a successful attack within their organization.However, even in a case like Astra, where the tool had found previously unknown vulnerabilities and built a full compromise chain that escaped the sandbox, the core challenge for defenders remains the same. The incidents security teams investigate will originate in familiar places, and adversaries will have similar goals to exploit exposed systems and steal credentials.AI-enabled attackers do not change what needs fixing for security teams. The difference is that defenders will have a thinner margin for error and the need to act at machine speed.Build for resilienceIt doesn’t makes sense to chase every new AI development with a “boil the ocean” security philosophy. Teams need to build a resilient security operation that continuously reduces exposure, swiftly detects threats, and systematically responds urgently when incidents occur. The question for security leaders in the next board meeting: How well can the organization sustain and mitigate attacks when our adversaries are armed with frontier AI model capabilities?The restricted access to Astra-level capabilities will not last forever. Capabilities spread and the safeguards put into place by frontier AI models are often aimed at protecting the provider’s platform, rather than the receiver of its output. As organizations aim to build a resilient strategy, keep five priorities in mind:
No single organization can prevent every attack originating from adversaries with access to frontier AI models. However, successfully secure organizations can withstand disruption, contain threats, and recover swiftly with minimal impact. Resilience will stand the test of time, regardless of AI-enabled attacker capabilities.Laura Ellis, senior vice president, AI, Arctic WolfSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
The company’s decision to restrict the most advanced cybersecurity capabilities under Astra only to those with access through OpenAI's Daybreak Blue should serve as a reminder of how powerful frontier AI models have become.Several autonomous, unsanctioned AI-driven attacks have played out in the public eye over the past several months, leading the industry to proceed with caution when releasing models like Astra. When a company determines that models are too capable to release to the public, security teams get a preview of the offensive power that will eventually reach adversaries with no restrictions.Race against the clockAs frontier AI models are shrinking the time between vulnerability discovery and exploitation to near-zero, security teams have struggled to keep pace. The challenge was evident in our recent report, where 70% of security leaders said an undetected threat has already resulted in a successful attack within their organization.However, even in a case like Astra, where the tool had found previously unknown vulnerabilities and built a full compromise chain that escaped the sandbox, the core challenge for defenders remains the same. The incidents security teams investigate will originate in familiar places, and adversaries will have similar goals to exploit exposed systems and steal credentials.AI-enabled attackers do not change what needs fixing for security teams. The difference is that defenders will have a thinner margin for error and the need to act at machine speed.Build for resilienceIt doesn’t makes sense to chase every new AI development with a “boil the ocean” security philosophy. Teams need to build a resilient security operation that continuously reduces exposure, swiftly detects threats, and systematically responds urgently when incidents occur. The question for security leaders in the next board meeting: How well can the organization sustain and mitigate attacks when our adversaries are armed with frontier AI model capabilities?The restricted access to Astra-level capabilities will not last forever. Capabilities spread and the safeguards put into place by frontier AI models are often aimed at protecting the provider’s platform, rather than the receiver of its output. As organizations aim to build a resilient strategy, keep five priorities in mind:
- Take inventory of everything internet-facing: Astra’s strongest results came against browsers and operating systems, so prioritizing internet-facing resources and practicing endpoint patch discipline deserves the same urgency security teams give to the perimeter.
- Identity represents the primary target: Deploy phishing-resistant multi-factor authentication and remove standing administrative privileges. Automated attacks still need a way in, and valid credentials remain the cheapest one.
- Compress detection and response: Faster attacks make decision latency the greatest organizational expense. Security teams should work toward 24x7 monitoring with built-in AI capabilities to isolate a host at 2 a.m. without requiring a meeting, and they may need to look to a third-party vendor rather than increasing their in-house oversight.
- Rehearse recovery: Run regular tabletop exercises in an AI-accelerated scenario, proving the organization’s ability to restore immutable backups inside a window security teams have measured.
- Govern the agents on the inside: Take inventory of the AI agents running in production, granting them least privilege access and requiring human review for consequential changes.
