At Syndis, we have been closely monitoring the latest AI developments, such as Anthropic's Mythos and Project Glasswing. This evolution confirms that artificial intelligence is vastly accelerating the discovery and exploitation of both known and unknown vulnerabilities (zero-days). However, it is crucial to separate the immense media hype from the actual, technical data available today.
Looking at public disclosures and CVE records, there are currently around 40 CVEs credited to Anthropic that are not related to Anthropic's own tools (like Claude Code). However, only one incident (CVE-2026-4747 in FreeBSD) is explicitly confirmed as autonomously discovered and exploited by Glasswing/Mythos; the other CVEs are credited to Anthropic researchers or partners. (Ref: https://www.vulncheck.com/blog/anthropic-glasswing-cves) This tells us that while the technology will undoubtedly be a game-changer, its real-world impact remains limited so far. The greater challenge lies in the vulnerabilities currently under embargo, which will be disclosed in Anthropic's comprehensive report in July 2026.
Below is an overview of how we at Syndis are addressing this development:
1. Monitoring and detection
Given that systems like Mythos will likely increase the volume of "zero-day" attacks, we place a strong emphasis on utilizing systems that rely on behavioral and machine learning analytics, such as EDR/XDR platforms. Furthermore, we are continuously integrating AI solutions into our own SOC processes to assist our analysts with deeper incident analysis. To ensure the utmost data security, this AI analysis runs locally on hardware owned by Syndis, rather than through external cloud services.
- Utilizing market-leading EDR/XDR solutions against AI threats: We utilize and manage robust solutions like Microsoft Defender XDR and Palo Alto Cortex XDR to defend against this trend. Even if the initial attack or vulnerability is discovered by AI like Mythos, these systems catch the consequences of the attack when exploitation is attempted on endpoints or within the network. They detect anomalous behavior, such as API key theft, abnormal code execution, and other lateral movement internally. This enables defenders to isolate and stop attacks in near real-time, regardless of whether the origin is artificial intelligence or a human.
Shadow AI: Just as we utilize logs in our SOC to detect anomalies, it is vital to monitor the use of unauthorized AI solutions within corporate environments. This allows security teams to detect traffic heading to these unapproved solutions and intervene to prevent unintentional or unauthorized data leaks.
Specific AI rules and "Prompt Injection": It is important to have realistic expectations regarding what data is actually accessible. Standard audit logs from Enterprise solutions like ChatGPT and Gemini generally do not include the actual input strings (prompts). To truly monitor and detect attacks like "prompt injection," you must look at the underlying architecture. This is especially true if you have proprietary software or web services where external users can input data that is then forwarded to an AI model. To defend against and monitor this, it is usually necessary to implement specialized AI security solutions (e.g., an AI Gateway or LLM Firewall) that sit between the software and the AI, and route the alerts from those systems into a SOC environment.
2. Threat intelligence
Developments like Glasswing undeniably impact Threat Intel processes, but we strongly emphasize driving these processes based on verified data and actual analysis rather than general media coverage.
- Faster vulnerability management and intelligence gathering: Since AI allows both researchers and attackers to find and write exploits in record time, we have fine-tuned our processes to capture intelligence on new vulnerabilities even faster. This enables us to promptly relay accurate information and mitigation recommendations.
Targeted warnings: If we observe campaigns utilizing AI-driven tools, we prioritize issuing specific warnings alongside actionable information on how to defend against them.
3. Incident response (IR)
While core incident response methodologies still apply, the unique nature of AI systems must be taken into account:
- Specific response plans (Playbooks): Our team of experts is continuously developing incident response playbooks to address emerging scenarios, such as the misuse of API keys for Large Language Models (LLMs) and data leaks involving AI services.
Speed of response: We place an even greater emphasis on rapid response to quickly isolate machines or user accounts the exact moment a security incident occurs.
4. Recommendations for your organization
To ensure your organization is prepared for the coming wave of new vulnerabilities and to support the secure use of AI across your business operations, we recommend taking the following steps:
- Hardened patch management in the coming weeks: Based on available data, Mythos appears particularly capable of finding vulnerabilities in older, established software. It is absolutely critical that your patch management processes are bulletproof, especially for internet-facing systems. We strongly recommend using the coming weeks and months to update all systems to their latest versions before the embargo on discovered vulnerabilities is lifted in July.
AI asset & integration inventory: The first step is to gain a clear overview. This includes not only mapping which external AI solutions are permitted within your organization but also identifying exactly where and how AI models have been integrated into your own systems and software development. Security teams need to know exactly which systems and processes they are protecting in order to provide targeted advice and effectively secure the infrastructure.
Identity and access monitoring: By ensuring that authentication logs (e.g., from your SSO systems) for logins to AI systems are ingested into your SOC or centralized logging environment, your security team can continuously monitor for suspicious activity. This allows for the verification of Multi-Factor Authentication (MFA) enforcement and rapid intervention if employee credentials are compromised.