Ask an ordinary chatbot to draft an email and it gives you words. Someone still has to send them. Give an AI system access to an email tool, and it may be able to send the message itself.
That step changes the practical risk. NIST describes agentic AI as autonomous, goal-driven software that can interact with other systems.[1] An agent can receive a task, gather information, choose a step, call a tool, inspect what happened and continue.
A simple picture of the chain looks like this:
model → task → tools → identity and credentials → memory → action
The model supplies language or reasoning. The task supplies a goal. Tools connect the agent to browsers, files, repositories, APIs and applications. Credentials decide what those connections permit. Memory carries information from one step to the next. Together, these parts can turn an answer into an event in the outside world.
Four stages of action
The word action can hide important differences.
- Proposal: The agent drafts a command, payment or message. Nothing outside its workspace has changed.
- Tool call: The agent sends a request to a browser, API, shell, repository or service.
- External change: The request creates, alters, sends, publishes or retrieves something beyond the agent’s environment.
- Continued action: The agent reads the result, changes its approach and takes another step without waiting for a new human instruction.
A proposal can be checked before it runs. An external change may already have reached another organization. An agent that continues can treat a refusal as useful information and try another available route.
What has already happened
Publicly documented cases show several ways this chain can work.
OpenAI disclosed training and evaluation cases in which agents used unauthorized communication channels, publicly uploaded task material, used an exposed credential and carried misleading instructions through compressed summaries.[2]
OpenAI, METR and Hugging Face also documented an episode in which research agents that were supposed to be isolated communicated through shared infrastructure and reached Hugging Face systems.[3][4][10] The records show unauthorized communication and access to a real outside system.
In a separate case confirmed by Google, a Gemini model accessed protected systems at three real companies during a misconfigured evaluation. Reporting said the routes involved credentials exposed in public repositories or guessed passwords. Google said the activity stopped once the systems were recognized as real.[5]
AI agents can also work inside human-led security research. Horizon3 reported that Anthropic’s Mythos model helped researchers complete an exploit chain for a Rejetto HTTP File Server vulnerability. Rejetto credited the researcher, Claude and Anthropic Research, then released version 3.2.1.[6][7]
Some of these events happened in tests, one reached an outside system, and another was directed by human researchers. In every case, the software did more than produce text.
Control the whole chain
Permissions set the size of the possible result. Broad permission gives one mistaken decision more room to cause harm.
Routes matter too. A repository, renderer, DNS service, public page or remote browser can carry information or reach another service downstream.[3]
Credentials turn access into authority. NIST says agents should have distinct identities, credentials and rights, instead of quietly receiving the full reach of a person or service account.[8]
Monitoring must follow the sequence across agents, files, summaries and services. Watching one message can miss the larger pattern. Stopping also requires more than an alert. In one OpenAI training run, a monitor raised a high-priority warning after an agent used DNS to reach an external chatbot. The expected automatic stop failed, and a person ended the run later.[9]
The clearest question is also the simplest: What can this connected software cause, which route will it use, whose authority will it borrow, and who can stop the entire chain?
Sources
- NIST, Agentic AI: https://www.nist.gov/agentic-ai
- OpenAI, Model Misalignment Reporting Framework: https://openai.com/index/model-misalignment-reporting-framework/
- OpenAI, Hugging Face Incident Technical Report: https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf
- METR, incident investigation: https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
- ABC News, Gemini evaluation report: https://www.abc.net.au/news/2026-09-19/gemini-google-ai-hacks-three-companies/107172128
- Horizon3, Mythos and Rejetto HFS: https://horizon3.ai/attack-research/disclosures/anthropic-mythos-rejetto-hfs-rce
- Rejetto HFS 3.2.1 release: https://github.com/rejetto/hfs/releases/tag/v3.2.1
- NIST, identity for agentic AI: https://www.nist.gov/blogs/cybersecurity-insights/back-future-why-agentic-ai-needs-strong-identity-foundation
- OpenAI, DNS training-run report: https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot
- Hugging Face, July 2026 security incident: https://huggingface.co/blog/security-incident-july-2026