Companies, governments and other organizations probably have less than six months to reinforce cybersecurity defenses for the coming age of frontier AI models.
This year we got a first-hand glimpse of the power of frontier AI models and the potential disruption they can cause in the hands of criminals, hostile governments and other threat actors. As Chief Information Security Officer for Broadridge, I oversee the internal team assigned to test a Mythos Preview model as part of Project Glasswing. Anthropic launched the initiative earlier this year to allow companies that maintain critical digital infrastructure to use the model to find vulnerabilities and shore up their security ahead of an inevitable wave of attacks when the model is released to the public.
By delaying the release of Mythos and giving the good guys advanced access, Anthropic helped us prepare for the upcoming increase in cyber activity. But that will be the case only if companies, governments and other organizations move quickly. In June, Anthropic predicted that frontier models from other providers would match Mythos capabilities in six to 12 months. That gives cybersecurity professionals less than a year to reinforce defenses before broad distribution of new frontier models puts this powerful technology into the hands of threat actors.
Companies in the expanding list of Glasswing participants have been working furiously to identify and eliminate specific vulnerabilities uncovered by Mythos. All other companies will have to quickly conduct that exercise when Mythos and equivalent models are released to the public, and every time a new frontier AI model hits the market.
But direct response to specific vulnerabilities will not be enough. There are many other steps organizations will have to take to protect against unknown vulnerabilities that AI will identify, create and exploit in the future. I provide a list of 15 of the most critical issues organizations will have to address.
While companies can rank these initiatives in terms of importance in the context of their own system configurations and businesses, ignoring or falling short on even just one could leave them dangerously compromised. Given the looming deadline, companies should be delegating these projects to separate teams in their cybersecurity departments to work on them in parallel, and with highest priority.
Monitoring and Visibility
1. Visibility expansion
In the age of frontier AI, companies cannot afford to have blind spots. At the very least, they should do everything they can to minimize external communications in which security tools can’t see or analyze the traffic. That includes inbound encrypted Transport Layer Security (TLS) traffic, which at this point should always be put through inspection. At a general level, companies can enhance visibility by ensuring that all security alerts are logged and onboarded to the Security Information and Event Management system (SIEM) and by implementing robust outbound traffic analysis. Finally, companies should re-assess any exceptions, bypasses or muted alerts on their Security Operations Centers (SOC). In many cases, it might make sense to subject these alerts to AI analysis as a means of double-checking.
2. AI for defense
Organizations should be using AI to get better at recognizing anomalies, patterns and deviations that suggest unusual activity. This should include applications that identify “AI-like” behavior due to speed or type of requests. Cybersecurity teams should also implement AI agents for SOC, triage, monitoring and threat hunting. Internally, organizations will have to manage fast-proliferating non-human identities (NIHs) in a way that that allows security teams to identify specific agents acting in their systems.
3. Application vulnerability
Frontier AI models will increasingly use valid credentials to target applications, for example leveraging data-brokers/info-stealers to gather valid credentials and successfully pass authentication checks. It is key that SOC teams ingest application logs to identify probing of internet facing applications, Frontier AI models will try to identify application logic weaknesses and this will likely create unusual patterns which standard cybersecurity tools will miss.
Perimeter Enhancements
4. Surface reduction
Eliminate any non-essential port, protocol or point of access to significantly reduce the risk of a frontier AI model finding and exploiting a vulnerability. Being thorough with non-essential is key.
5. Segmentation
Take the “assume breach” maxim seriously. Implement micro-segmentation controls sufficient to ensure survival when one or several devices are compromised. Pay special attention to dependencies with core systems like Active Directory and DNS, where AI would still have access even after segmenting the main servers.
6. Restrictions
AI models are accessing manuals online and exploiting the massive SaaS footprint to gain access. That makes foundational tools like Multi-Factor Authentication (MFA) Single Sign-On (SSO) and comprehensive IP restrictions critical. As the risk of command-and-control (C2) back channel and data exfiltration increases, companies must enhance outbound restrictions and sinkholes that can intercept outbound traffic requests.
Internal Network and End User Protection
7. Inbound protections
Inbound email filters have improved dramatically in recent years, with AI now being used to assess language and tone to identify phishing.
8. Desktop/endpoint restrictions
Employees might not like tough restrictions on software installation, and tight controls on execution, script execution and PowerShell can cause some minor inconveniences for staff. But these essential precautions will reduce the speed of execution, raise alerts and hopefully make sure the good guys have time to respond before damage is done when a device is compromised.
9. Remote Browser Isolation (RBI)
Security professionals have long wrestled with the tradeoffs associated with RBI, which helps protect networks but can also introduce latency and slow and degrade Internet performance for users. With Mythos and other frontier AI models now in the mix, organizations can no longer do without this effective barrier to phishing and fake websites, especially when RBI technology is getting better every day.
Vulnerability Management
10. Vulnerability remediation
Mythos and other frontier AI models will give threat actors enhanced ability to chain together low- and medium-level vulnerabilities to access systems. Organizations need to conduct a comprehensive review of their networks and security policies to account for increasing exploitability.
11. End of Life (EOL) and End of Support (EOS) removal
Given threat actors’ enhanced speed and capabilities, companies must act to immediately retire, replace, isolate, or apply compensating controls to exposed EOL and EOS assets. Any system that can’t be patched quickly now represents an unacceptable risk.
12. Appliance monitoring
Network appliances, which often sit in central locations, can’t run the End-Point Detection and Response (EDR) tools and vulnerability-management agents generally used to monitor internal processes and flag concerns. That makes them a risk. That risk can escalate rapidly if a threat actor using a frontier AI model identifies a vulnerable port or obtains credentials. Companies need to update their policies regarding appliance use and significantly step up monitoring.
13. Privilege management
Frontier AI models make it easier for threat actors to hijack privilege access — including those of both human users and internal non-human AI agents. Because external AI models can use valid identities to access applications and infrastructure, managing privilege access accounts has become more important than ever. This should include regular recertification and prompt removal when an employee leaves the company, changes roles, or otherwise no longer requires the privileged access.
14. Server security
In the age of powerful frontier AI models, all companies should prioritize enhancing fundamental server security precautions such as authentication, access controls, patching, vulnerability management, firewalls, EDR, logging and monitoring, encryption, backups, and configuration management. Companies should also go the next step and move to harden all servers by eliminating unnecessary software, services, open ports, administrator accounts and other permissions. In a best-case scenario, companies will adopt hardware keys to dramatically minimize the risk of remote attacks.
15. Response preparation
Ensure the organization has sufficient automation to mount an autonomous response to what could be a fast increase in the number of AI-driven attacks. Adopt pre-approved execution plans for containment.
Addressing the Issues
If necessary, companies should start with a few mission-critical enhancements from this list. But ultimately, companies will need to address every one of these issues and more to be ready for the rapidly approaching day when frontier AI models are broadly available to threat actors.