Vulnerability Discovery and Exploitation Trends in the AI Era

MITRE Techniques

• [T1190] Exploit Public-Facing Application – Threat actors targeted exposed enterprise and AI middleware services and unauthenticated endpoints for initial access; cited as exploiting public management interfaces and exposed middleware (‘unauthenticated public management interfaces’, ‘exposed middleware’).
• [T1068] Exploitation for Privilege Escalation – Post-exploitation activity included privilege escalation after initial compromise of vulnerabilities such as CVE-2026-1731 (‘post-exploitation activities, including privilege escalation’).
• [T1059] Command and Scripting Interpreter – Several flaws enabled command injection or Python code execution, allowing adversaries to run commands on the host (‘OS command injection’, ‘Python code injection via exec()’).
• [T1195] Supply Chain Compromise – The report describes vulnerabilities in software ecosystems, vendor disclosure cycles, and AI/LLM stack components that can be abused before downstream deployment (‘software or services to other enterprises and consumers’).
• [T1021] Remote Services – Threat activity involved compromise of remote support and privileged access services, enabling remote interaction with targeted systems (‘BeyondTrust Privileged Remote Access (PRA) and Remote Support’).
• [T1005] Data from Local System – Attackers were observed stealing API credentials and exfiltrating data from compromised systems (‘API credential theft’, ‘data exfiltration’).
• [T1074] Data Staged – The report notes theft of credentials and private prompt streams from AI gateways, implying collection of sensitive data before exfiltration (‘private prompt streams containing personally identifiable information’).
• [T1105] Ingress Tool Transfer – Secondary payloads such as SNOWLIGHT and SPARKRAT were dropped onto compromised systems (‘dropping secondary payloads including SNOWLIGHT, SPARKRAT’).
• [T1053] Scheduled Task/Job – One Langflow issue allowed remote actors to drop cron jobs onto hosts, indicating scheduled execution persistence (‘drop unauthorized files (e.g., cron jobs)’).
• [T1106] Native API – Exploitation of unauthenticated administrative APIs and AI platform endpoints was used as a direct entry path (‘Unauthenticated Administrative APIs’).
• [T1211] Exploitation for Defense Evasion – AI gateways and edge appliances were abused to bypass perimeter controls and EDR blind spots (‘bypass perimeter firewalls’, ‘enterprise EDR agent blind spots’).