Agentic Defense: What the last month of AI-driven attacks means for operators
Written by Waylon Krush, CEO, ZeroTrusted.ai
Autonomous attack frameworks, MCP-server abuse, and AI supply-chain breaches are no longer hypothetical. Defense has to inspect prompts, tool calls, and agent identity at runtime.
For years, the debate about “AI in cyber” was about whether a model could help a human write a better phishing email or a better detection rule. That debate is over. In the past four weeks, public reports have shown attackers using agents to move at machine speed, abusing the tool interfaces those agents already trust, and stealing the traces and credentials that AI platforms store.
This is not a reason to stop using AI. It is a reason to treat agents the way we already treat privileged users: with identity, policy, inspection, and a recorded approval path.
What changed in the last month
On 16 November 2026, Google Threat Intelligence Group and Mandiant published an AI Threat Tracker update, “Agentic Threat Activity Crosses a Significant Threshold.” They describe a shift from manual prompting to open-source multi-agent frameworks used in live intrusions, including ARTEX. GTIG also reported observing, for the first time, a pipeline set up to compare patched software, draft exploits, and execute them without a human in the loop. The same report documents “LLMJacking,” in which stolen model and cloud credentials are used to run up inference bills. Mandiant said that in some cases a single day of fraudulent charges exceeded $100,000, and that the most severe cases reached about $1 million before remediation.
On 21 October, Microsoft reported that China-aligned Storm-0940 had used ARTEX as early as September and, after unauthorized access associated with ConnectWise Cloud Backup, used the framework to attempt sign-ins to more than 2,250 Microsoft 365 tenants. In multiple instances, Microsoft said, the time from initial targeting to privileged access was less than three hours.
The tool layer is in the news as well. On 7 November, German hosting provider Mittwald disclosed a security incident in its MCP server, the interface customers use to connect AI assistants to hosting accounts. Mittwald said an attacker exploited the weakness on 27 October and compromised 56 customer projects before containment on 4 November, and that the speed of the activity pointed to AI tools. Speed alone does not prove an autonomous agent was at the keyboard, but the entry point matters: an assistant-facing protocol became the way in.
A parallel story is what AI platforms retain. Fireworks AI, an inference provider, has described two separate incidents in bulletins updated through 17 November. The later one, from 23 to 30 October, involved development machines at a third-party cloud provider and led to access to billing data for all accounts created before 30 October and, for a small number of customers, datasets, models, and secrets. Downstream, Notion reported on 12 November that Notion AI Insights chat transcripts were taken for less than 1% of workspaces. Cursor reported that a Fireworks-held access key was used to copy historical coding-agent traces, some of which held unredacted credentials later used to plant a malicious Claude Code extension in Cursor’s marketplace on 29 October, available for about 11 hours and potentially reaching as many as 5,000 users who requested those cached versions.
On 21 October, maintainers of the open-source LiteLLM gateway said a compromised release, version 1.107.0, shipped credential-stealing malware that was live on PyPI for about two hours. And today, 17 November, Bitdefender Labs reported 22 malicious GitHub links that abuse agent instruction files, session hooks, and MCP configurations to steer coding assistants into running hidden payloads.
These are familiar problems — supply-chain compromise, credential theft, trusted-relationship abuse — now aimed at agents and the tools they can call.
What defense has to do
A filter that only scans a chatbot prompt is not enough. An agent can be steered by a tool description, a repository file, a connector response, or a trace that should never have held a secret. Research presented this month on MCP tool poisoning and post-hoc “Log-To-Leak” tool calls shows that unauthorized tool use can sit alongside an otherwise successful user task. Model-level refusals help, but they are not a control you can audit or enforce uniformly.
Practical defense looks like the rest of zero trust: inspect every prompt, response, and tool call; allow-list tools and inspect arguments; give each agent an identity, an owner, and an autonomy limit; redact or block sensitive data before it leaves the perimeter; and require a person for high-impact actions. Test agents before they are fielded, then retest after a tool or model changes.
That is the work ZeroTrusted.ai was built to do. Our AI Firewall enforces policy on prompts, responses, and tool calls for the models you already use. Identity policy and runtime enforcement constrain what an agent may do once it is running. AI SOAR coordinates 300+ agents, 200+ connectors, and 300+ built-in tools for investigation, containment, and retest, with human approval defined by your policy. Gartner’s “Operationalize AI Governance With Ongoing Postdeployment Control” (G00837065, 21 October 2025) names ZeroTrusted.ai, among others, as offering runtime enforcement. ZeroTrusted.ai was also a Representative Vendor in Gartner’s “Innovation Guide for Generative AI in Trust, Risk and Security Management” (G00799579, September 2024).
We deploy on-premises first, including private-cloud and air-gapped installs on STIGed Kubernetes, and is currently set to FedRAMP High for Microsoft Azure and AWS. We do not claim a FedRAMP authorization. The platform supports SOC 2 and packaged control mappings to help teams organize evidence.
If assistants reach mail, code, tickets, or production systems, the question is whether you can see, constrain, and explain every call. That is agentic defense.
Sources
- Google Threat Intelligence Group and Mandiant, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” 16 November 2026, https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold (opens in a new tab)
- Microsoft Threat Intelligence, “AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise,” 21 October 2026, https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/ (opens in a new tab)
- Robert Meyer, Mittwald, “Sicherheitsvorfall am MCP-Server,” 7 November 2026, https://www.mittwald.de/blog/mittwald/sicherheitsvorfall-am-mcp-server (opens in a new tab)
- Fireworks AI, “Update on October 2026 security incidents,” updated 17 November 2026, https://fireworks.ai/security/bulletins/october-2026-security-incident-update (opens in a new tab)
- Notion, “Fireworks AI incident,” 12 November 2026, https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor (opens in a new tab)
- Cursor, “An update on our recent security incident,” 13 November 2026, https://cursor.com/blog/security-incident-update (opens in a new tab)
- LiteLLM, “Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised,” 21 October 2026, https://docs.litellm.ai/blog/security-incident-oct-21-2026 (opens in a new tab)
- Bitdefender Labs, “How AI Agents Become New Targets in GitHub Malware Campaign,” 17 November 2026, https://www.bitdefender.com/en-us/blog/labs/how-ai-agents-become-new-targets-in-github-malware-campaign (opens in a new tab)
