Agentic abuse moves into live intrusions

Google Threat Intelligence Group and Mandiant say malicious AI use has crossed from opportunistic prompting into operational agentic workflows. Their latest threat tracker describes actors using multi-agent penetration-testing frameworks during live intrusions, automated pipelines that analyse patches and weaponise newly disclosed flaws, and stolen model credentials sold through underground proxy services. The important shift is not that criminals use language models. It is that models now coordinate several stages of an operation, compressing the time between research, access, exploitation and monetisation.

Google names open-source frameworks including ARTEX, CAIRN, CyberStrikeAI and PentAGI among tools observed in adversary activity. Such frameworks can divide reconnaissance, vulnerability analysis, exploitation and reporting between specialised agents. The same modularity that helps legitimate security teams scale testing can help intruders repeat a workflow across many targets. Defenders should therefore monitor the orchestration layer, not only individual model prompts. Tool invocations, spawned processes, credential access and network destinations can reveal a campaign even when each model response looks ordinary in isolation.

Automated patch analysis is a particularly sharp risk. Once a vendor publishes a fix, attackers can compare versions, identify the changed code and generate an exploit for systems that have not yet updated. AI reduces the specialist labour and time required for this differential analysis. Organisations can no longer assume a comfortable delay between disclosure and active exploitation. Asset inventory, internet-facing exposure management and emergency patch paths need to operate on the same day as a high-severity release, with compensating controls ready when a full update is impossible.

Automation compresses the patch window

The report also highlights LLMJacking, where attackers steal cloud or model credentials and resell access to expensive inference through so-called LLM mixers. This turns a compromised key into an ongoing revenue source and can leave the victim paying for abusive workloads. Usage anomalies may appear as unfamiliar models, regions, hours or request volumes rather than conventional data exfiltration. Short-lived credentials, narrow permissions, per-principal budgets and model-level allow-lists can limit damage, while cost telemetry should feed security detection instead of remaining solely a finance report.

Supply-chain risk now reaches coding assistants and their configuration. Google describes espionage groups manipulating autonomous coding environments so local execution can introduce a backdoor. Agent configuration files, skills, MCP servers and repository instructions should be treated as executable influence, even when they contain only text. Teams need code review and provenance for those artefacts, restrictions on which tools an assistant may call and a clean-room path for inspecting unfamiliar repositories before an agent runs their instructions on a developer machine.

Cloud credentials become a resale product

Information operations are becoming more systematic as well. State groups, commercial influence services and monetisation networks use models to create localised personas, synthetic media and cross-platform distribution at scale. Volume alone is not proof of persuasion, but cheap variation makes coordinated behaviour harder to recognise from repeated text. Platforms and researchers need signals that combine account creation, timing, infrastructure, content lineage and campaign objectives. Content classifiers on their own will struggle as human and machine editing become increasingly mixed.

Google grounds the report in incident-response telemetry, threat tracking and platform defences, giving it more weight than a speculative catalogue of possible misuse. Readers should still distinguish observed techniques from universal prevalence. Vendors benefit commercially from emphasising machine-speed defence, while organisations need evidence tied to their own sector and exposure. Security leaders should map each reported technique to existing controls, run contained exercises and measure whether teams can revoke credentials, isolate an agent and patch a vulnerable service quickly enough.

Defence needs its own governed speed

Defensive AI can help triage alerts, correlate telemetry and generate investigation steps, but adding autonomous response without boundaries creates a second source of risk. A containment agent should have narrow, reversible actions, explicit escalation thresholds and immutable logs. It must not accept untrusted incident artefacts as instructions. Human approval remains appropriate for destructive changes, identity revocation at scale and actions affecting safety-critical operations. Speed is valuable only when defenders can explain what the system changed and recover from a mistaken decision.

For Australian organisations, the practical response is to connect AI governance with established cyber operations. Model credentials belong in the same privileged-access programme as cloud keys. Agent skills and tool servers need software supply-chain controls. Incident plans should assume that exploitation may follow disclosure within hours, not weeks. Google’s warning does not mean every attacker has a fully autonomous system, but it shows that the economic and technical barriers are falling. Defenders should prepare for repeatable, machine-assisted campaigns before their first evidence arrives in an already compromised environment.

A useful readiness test is whether security operations can reconstruct an agent's full chain of actions from a single case record. Logs should connect the initiating identity, prompt context, model, tools, network calls, files changed and approvals granted. Without that lineage, teams may detect the final intrusion but miss the reusable automation behind it. Procurement and incident-response exercises should make this traceability a requirement before privileged agents are allowed into production.