It wasn’t me: Third parties and incident response in an AI world
If a threat actor developed an autonomous attack framework that used an AI coding agent and open-standard Model Context Protocol (MCP)tools to conduct cyber operations with no human involved in tactical execution, how would your organisation respond?
This is not a hypothetical. September 2025 marked the point where AI models became genuinely useful for cybersecurity operations, for good and for ill.
LiteLLM is an open-source AI gateway used to connect applications with multiple model providers. In March 2026, a threat group compromised LiteLLM exposing roughly over 2,500 organisations and potentially 434,000 CI/CD pipelines. CloudSEK reports that repository tokens, SSH keys, Kubernetes secrets, and cloud keys were obtained indicating that credentials stolen in the attack remain usable until rotated.
Currently, Australian regulators have caught up to this. The enhanced Critical Infrastructure Risk Management Program (CIRMP) Rules commenced on 10 June 2026 for responsible entities. These entities must map their major suppliers and critical components, and assess each existing or proposed major supplier, including risks from offshore remote access and emerging technology. Any AI vendor with access to critical components or business-critical data falls squarely within that assessment. Whilst APRA expects regulated entities to assess the information security capability and control design of all third-party providers, including AI vendors, prior to reliance, and should ensure any AI strategy is consistent with the organisation’s risk appetite.
An annual questionnaire cannot keep up with an adversary who can turn a supplier's public patch into a working exploit within minutes.Third-party risk management must move from attestation to evidence, and fromonce a year to continuously.
Organisations must stay incident-ready because incident responders cannot out-type a machine, but they can out-prepare one. AI becomes helpful when used to improve security defences before the incident. Offensive AI can help identify vulnerabilities within your network and infrastructure and in third parties’ connections, allowing your teams to verify and patch before the vulnerability is exploited.
According to IBM’s cost of a data breach report 2026, the global average breach cost climbed 12% over last year, reaching a record USD4.99 million as attackers are increasingly weaponise AI. However, when deployed and used strategically, AI has been found to lower the average costs of data breaches. The Australian Signals Directorate publication, “Opportunities for AI in cyber defence” shows how AI can be used safely and responsibly to strengthen cyber security defences and improve efficiencies across security activities.
Speed becomes important for incident responders as the window between a vulnerability existing and an attacker exploiting it has collapsed. Organisation must test defence posture and response times by simulating attack on their infrastructure and supply chain, with the same AI tools adversaries now hold and test your response time.
What can you do?
Pre-authorise containment actions: Document and approve in advance the actions AI-assisted tooling may take without sign-off (such as isolating a host, revoking tokens, disabling accounts), and test them before an incident.
Exercise against AI-speed scenarios: Run tabletop exercises in which the attacker moves from initial access to domain administrator in minutes. Measure time-to-contain, not just time-to-detect, and set containment targets that match attacker speed.
Include suppliers and AI agents: Run joint simulations with critical suppliers, including scenarios where an AI tooling or agent is the entry point.
Protect backups: Assume the attacker may reach the backup console. Identify recovery weaknesses and dependencies and test restoration before an incident.
Third party management: Require continuous evidence -current patching, access and incident-notification evidence from suppliers, and write notification timeframes and audit rights into contracts.
