Governed AI Can Speed Security Response, but Human Oversight Must Set Limits

Summary
A Securonix field CISO argues that governed AI can speed security investigations and response, but teams should set permissions in advance, preserve evidence, test controls, and require human approval for potentially disruptive actions.
Key points
- ITRC recorded 3,064 publicly reported data compromises by September 30, 2026; it also found that 77% of Q3 public breach notices omitted an attack vector, limiting confident claims about AI’s role.
- IBM reported an average breach cost of $4.99 million and said AI-enabled malicious breaches averaged about $6 million; the article cautions that the association does not establish how much AI alone added to costs.
- The article recommends defining AI’s permitted actions before incidents, enforcing limits outside the model, preserving investigation evidence, and requiring human approval for actions that could disrupt critical services.
- The PeopleSoft example shows how a web application firewall rule matching a literal path could miss an encoded request that the application later decoded; the author recommends testing controls as requests are processed and patching exposed software.
- Stolen OAuth tokens and session cookies can let attackers use legitimate cloud APIs and export functions, so investigators should compare activity with audit records and business context before concluding it is malicious.
- For AI agents, the article recommends limiting privileges, authorizing individual tool actions, and keeping audit context sufficient to investigate actions outside the agent’s assigned work.
- The author advises evaluating security providers with realistic scenarios and checking whether their systems actually block access, preserve evidence, and allow analysts to reverse mistaken actions.
Article Details
- Topic
- Governed AI and human oversight for faster breach investigation and response, amid AI-enabled attacks and persistent access risks.
MITRE ATT&CK
T1053.005 · Scheduled TaskTASK#STOMP uses scheduled tasks as a persistence mechanism.T1059.001 · PowerShellTASK#STOMP uses PowerShell as part of its operation.T1190 · Exploit Public-Facing ApplicationAttackers exploited a vulnerable PeopleSoft component and PHP flaws to gain remote code execution.T1528 · Steal Application Access TokenThe article describes theft of OAuth and integration tokens to access enterprise services.T1566.004 · Spearphishing VoiceVoice phishing and help-desk manipulation were described as methods for obtaining access.
CVE
Threat Actors
ShinyHuntersThe article describes UNC6240, which was identified in renewed PeopleSoft exploitation, as associated with ShinyHunters; it does not independently attribute that activity to ShinyHunters.UNC6240Mandiant and Google Threat Intelligence Group identified renewed exploitation of PeopleSoft by UNC6240, including requests that bypassed a web application firewall rule.
Malware
Vendors
InstructureInstructure’s explanation of the Canvas incident described another route to elevated access: an attacker submitted malicious code in a support ticket, and a cross-site scripting vulnerability allowed an authorizationMicrosoftFor example, Microsoft notes that Entra ID cannot directly revoke a session token issued by an application, because the application controls that session.OracleOracle described CVE-2026-35273 as remotely exploitable without authentication, potentially allowing remote code execution.
Products
CanvasInstructure’s explanation of the Canvas incident described another route to elevated access: an attacker submitted malicious code in a support ticket, and a cross-site scripting vulnerability allowed an authorizationEntra IDFor example, Microsoft notes that Entra ID cannot directly revoke a session token issued by an application, because the application controls that session.PeopleSoftThe PeopleSoft attacks documented this year show a software exposure surviving an apparent mitigation.PHPThe October PHP findings put a concrete example behind existing weaknesses.
