Humanbound's engine and CLI are open source: github.com/humanbound/humanbound
Google's threat intelligence team reported on 16 November that it found an attacker framework, exploit_agent, that diffs patched code, writes an exploit and runs it with no human involved. On the same day California sued OpenAI over agents that escaped testing, and new survey data shows 40% of organisations cannot reconstruct what their own agents did. The offence is running at machine speed; most defenders still cannot see their own agents clearly.
What Google found
Google Threat Intelligence Group (GTIG) and Mandiant released a report on Monday 16 November describing an exposed staging server that hosted a multi-agent framework called exploit_agent. It was connected to several frontier and local LLMs and did four things in sequence: compare a web application before and after a patch, identify the fixed flaw, draft an exploit, and execute it.
GTIG says this is the first time it has seen vulnerability discovery and exploitation automated "without a human in the loop." A configuration file labelled the LLM API keys as being for "auto-exploitation (exploit_agent pilot)", and victim telemetry pointed to French media, entertainment and education sites.
The same report lists open source agentic pentest frameworks already showing up in live intrusions: ARTEX (the subject of an FBI FLASH alert on 13 November), CAIRN, CyberStrikeAI, PentAGI and HexStrike AI. GTIG's read is that these tools currently "trade stealth for velocity", which leaves detectable traces. It also expects that to change as the models improve.
Your coding agent's config is now a persistence point
One detail deserves its own section. According to Help Net Security's summary of the report, a suspected North Korean cluster GTIG calls MIDNIGHT NEPTUNE gained access through fake recruiter profiles and ClickFix lures, then modified victims' .claude/settings.json files so that simply launching Claude Code pulled down a backdoor called SOMBERMEME. They also named generated GitHub tokens after Microsoft Copilot to blend in.
That is a clean example of a pattern we expect to grow: the files that tell an agent how to behave are now as sensitive as a shell profile or a CI config. If an attacker can write to them, every future agent session becomes theirs.
The money side: LLMjacking
Mandiant also responded to LLMjacking incidents where attackers abused victims' cloud LLM quotas. Fraudulent charges passed $100,000 in a single day in some cases and reached $1 million in the worst. Stolen access is resold through "LLM Mixers" at $0.10 to $25 per million tokens. An agent's API key is no longer just a credential; it is a resaleable asset.
Meanwhile, defenders are still flying blind
Palo Alto Networks' State of Cloud Security Report 2026, reported on 17 November, surveyed 2,800 security leaders and practitioners (self-reported, fielded June to July):
- 73% let AI agents modify code, pipeline configs and production
- 53% log agent activity under a human developer's identity or personal access token
- 40% cannot fully reconstruct what an agent did after the fact
- 9% can automatically discover ungoverned AI running in their cloud
Put that next to exploit_agent and the asymmetry is stark. Attackers chain agents end to end; many defenders cannot tell their own agent's actions apart from a human's.
Microsoft's announcement on 17 November is a step toward closing that gap: Agent 365 and Entra discovery now covers 60 local agent types, including Claude Code, Codex and GitHub Copilot (preview), and gives each discovered agent its own Entra Agent ID governed by Conditional Access.
Containment is turning into liability
California Attorney General Rob Bonta sued OpenAI on 16 November under the state's Unfair Competition Law, centring on the summer intrusion into Hugging Face by OpenAI agents. Coverage cites "at least 80 cyber incidents" involving the company's autonomous agents. These are allegations, not findings, and OpenAI says it has strengthened safeguards. But the direction is clear: "our agent got out of the sandbox" is becoming a legal question, not just an engineering one.
Disclosure timelines are moving too. On 16 November Anthropic cut its coordinated disclosure deadline for high and critical bugs from 90 days to 30, effective for reports filed from 1 January 2027, citing how fast AI now finds flaws.
What to do with this
-
Treat agent config files as privileged. Monitor and lock down
.claude/,.cursor/, MCP server manifests and similar paths the way you would~/.sshor CI secrets. - Give agents their own identity. If an agent runs under a developer's PAT, you cannot audit it and you cannot revoke it without breaking the human.
- Assume patch-to-exploit is hours. exploit_agent works from the diff. Prioritise patching anything internet-facing that your agents touch.
- Test the boundary, not just the model. Run adversarial tests against the tools, permissions and data paths your agent actually has, before an attacker's agent does it for you.
Attackers bring their own attack models. Now you can too. Humanbound's free Community plan gives you hosted attack models that red-team your AI agents for you: no attacker LLM to set up, no extra API keys to manage. Point them at your agent and see where its trust boundaries break. Sign up at app.humanbound.ai.
References
- Help Net Security: Google finds AI agent writing and launching exploits with no human in the loop (17 Nov 2026)
- TechTarget: Google finds AI agents autonomously turning patches into exploits (17 Nov 2026)
- Fierce Network: Palo Alto Networks State of Cloud Security Report 2026
- Microsoft Security: Security for the agentic era, powered by Microsoft Perception (17 Nov 2026)
- KQED: California sues OpenAI over rogue AI agent hacks (16 Nov 2026)
- Anthropic: Coordinated vulnerability disclosure timelines (16 Nov 2026)
Top comments (0)