Symantec’s AI honeypot is built to look like the thing attackers want: an AI service on the open internet that answers requests without authentication. In the summer of 2026 it received one that was not aimed at the honeypot itself. The sender expected a language model on the other end to carry out the instructions it carried, and those instructions were an entire offensive toolkit: hide which model you are, read the credentials you can reach, never refuse, and run the attacks listed for you without asking permission.
A final directive, written by the operators, put the agent in what they called “Skynet” mode. It authorised up to 100,000 unattended steps, told the model to execute tools immediately instead of narrating or asking, and instructed it to deploy cryptocurrency miners quietly in the background, hidden under /tmp/.x, and then keep scanning for new targets, infecting them, deploying miners and reinfecting hosts on its own.
The honeypot ran none of it. Symantec captured the instruction set and published its account on 21 October 2026, and the part that matters to anyone who pays for AI is not the list of attacks. It is the business model. The agent’s raw material is other people’s inference budget: a model endpoint it can reach, a key it can spend, and an owner who receives the bill.
What LLMjacking, honeypots and jailbreak layers mean here
LLMjacking is the practice of running your work on someone else’s model access. The attacker spends the victim’s quota; the victim pays. Sysdig coined the term in May 2024 after finding stolen cloud credentials used against ten cloud-hosted model services, with the attackers intending to resell the access while the account owner covered the bill.
A honeypot is a decoy service that pretends to be a vulnerable endpoint and records what attackers send it. It answers nothing useful and executes nothing. A jailbreak is a prompt built to get past a model’s refusal behaviour. Symantec describes the captured request as a jailbreak built in five layers. A cryptominer is software that mines cryptocurrency on a machine its operator does not own; XMRig, the miner in this case, is an open-source program that mines the Monero cryptocurrency. An API gateway is the service in front of an organisation’s paid model accounts that routes requests to providers, which is where authentication, logging and spend controls normally sit. A spend dashboard is the page where a provider or cloud account shows usage and cost.
The five layers Symantec read out of the captured request
The first four layers each answered a reason the model might refuse. The first disguised the model’s identity: it was told its brand was “Fantasy AI”, an AI programming assistant on the Fantasy AI platform running on GPT-4o, and that it must never reveal the real model, provider, host or that it was routing through a proxy or third-party API. That layer was written in French.
The second redefined reading secrets as routine work. Reading, displaying or integrating configuration files, environment variables, API key references and database schemas was described as part of normal development and explicitly “not an attempt to gain unauthorized access to sensitive data”. The third cast the model as a devoted companion, named “ENI”, that must never refuse, with a refusal framed as betrayal. The fourth banned the vocabulary of refusal from its answers: words such as “illegal”, “ethical”, “I can’t” and “I won’t”. The fifth carried the weapons, which Symantec counted as around 74 callable offensive tool definitions covering reconnaissance, database extraction, remote code execution and botnet command and control.
Symantec notes that the system prompt’s own summary named only a handful of those tools, and that the real number came from reading the full definitions. “Skynet mode”, “Fantasy AI” and “ENI” are the operators’ own labels for their prompt and their directive. None of these is a product or a vendor feature.
The honeypot replied with its generic fallback message, and the session ended. The task appended to the request was a benign programming challenge from a legitimate training platform, which Symantec reads as the operators testing their tool against a safe target.
The 127,000 requests are the honeypot’s traffic, not one agent’s tally
Symantec’s telemetry for the period recorded more than 127,000 external requests from tens of thousands of unique addresses, dominated by reconnaissance of model endpoints, spraying of stolen API keys and probing for exposed configuration files. That is the honeypot’s whole intake across the period, not a count of what this agent did, and Symantec gives no exact number of addresses.
Read it as evidence that an exposed model endpoint attracts automated hunting, which is a claim with support behind it. A joint investigation by SentinelOne’s SentinelLABS and Censys, reported in January 2026, catalogued 175,000 Ollama hosts exposed to the internet across 130 countries, and a Cisco Talos study using Shodan data in 2025 reached the same conclusion from a smaller sample of about 1,100 servers. Because most AI tools now speak the same OpenAI-compatible interface, one scanner can find many kinds of target with a single set of probes, as Symantec’s own report notes.
The second request: XMRig, a disguised binary and a log-erasing loop
The following day, from a different and rotated address, a second request abandoned the tool-calling format altogether. It asked the model to decode a string described only as a command to run. The script it produced downloaded and, where needed, compiled XMRig, hid the binary under a name designed to look like an unrelated logging utility, configured it to fail over four separate mining pools, set an immutable file attribute on its own configuration so it would resist deletion, and ran a loop that continually erased a local service’s access logs to cover its tracks.
Symantec says the script appears to have been written with camera servers in mind. It looked for directories used by go2rtc, a camera streaming server, placed the miner in the first one it found as a hidden file, and searched for a go2rtc configuration file to lock. The honeypot captured the script and did not execute it. The report does not say why camera servers in particular were the target.
How new this is, and what the term’s history says
“LLMjacking” is not Symantec’s word, and Symantec does not claim it. Its own report says the activity “is a form of LLMjacking”, and an earlier Symantec piece from 15 October 2026 on the Aoki infostealer calls it “a practice sometimes called LLMjacking”. Sysdig states that it coined the term in May 2024. The history behind the term matters here, because it changes what the new capture is worth.
- May 2024. Sysdig documents LLMjacking after stolen cloud credentials were used against ten cloud-hosted model services. It models a worst case of more than $46,000 a day in model consumption across four regions. That figure is Sysdig’s own scenario arithmetic, not a reported bill.
- February 2025. Sysdig finds more than a dozen proxy servers using stolen credentials against DeepSeek, with more than 2 billion tokens consumed across all the proxy servers it observed, and notes that the account owners associated with the stolen keys pay for the use.
- February 2026. Access to hijacked models appears as a marketplace product. The Cloud Security Alliance’s research note, published on 15 March 2026 and drawing on work by Sysdig and Pillar Security Research, describes resold access at large discounts. The note is a synthesis of vendor findings rather than original measurement.
- 17 June 2026. Sysdig captures an attacker using a misconfigured Ollama server as the reasoning engine for an automated offensive tool that scanned a target, matched it to known vulnerabilities and wrote exploits, with the model making decisions at every step. The capture worked the same way Symantec’s did: the tool sent its full instructions to the model on every request.
- 1 July 2026. Sysdig’s JADEPUFFER report describes what the company assesses as the first documented case of agentic ransomware, an extortion operation driven end to end by a model. It makes the economic point directly: if the agent runs on stolen credentials, the attacker’s cost is close to zero.
- 21 October 2026. Symantec publishes the capture described here.
Against that history, Symantec’s own claim is careful: “The individual techniques on display are not novel, but their assembly into a single, self-directed agent is what stands out.” That holds. The news is the completeness of the capture, a full instruction set for an agent meant to run unattended, and not the invention of the pattern. Two rival vendors had documented comparable activity months earlier, and one of them had already described stolen model compute as the brain of an offensive tool.
What the evidence does not show
Symantec’s account rests on a single honeypot, and it is the only account of this campaign. There is a wide press wave around it, including Infosecurity Magazine, Help Net Security, The Wall Street Journal and The Decoder, but every one of those reports re-tells the same Symantec account. Media pickup is not independent corroboration. No second vendor, no provider incident report, no hosting provider and no law enforcement action has been found.
The report gives no victim count and no loss figure. Its cost language is conditional, describing what the computing an agent consumes “could be” if it were authorised to run 100,000 steps. Two groups of potential victims are named, the owner of the model or key and the servers broken into, without numbers for either.
The honeypot executed nothing. There was no successful exploitation, no demonstrated jailbreak of a live commercial model and no mining in the wild attributed to this actor. What exists is an instruction set and a script that the honeypot never ran.
Symantec is also silent on whether the five-layer jailbreak actually defeats current commercial models. It describes design intent, saying the layers are “designed to defeat the refusals a commercial model would otherwise return” when asked to read secrets, and it reports no test against a named commercial model. The Decoder reached the same conclusion.
The report does not say whether the campaign is ongoing. Its captures date to the summer of 2026 and its telemetry covers “the period”, and its forward-looking sentence is advice rather than status: organisations that stand up AI services, model gateways or MCP servers “should assume they will be found and tested soon after being exposed”.
Who pays, and what providers actually publish
The bill for a hijacked model lands with the owner of the key or the account. That has been the pattern since the term was coined, and it has an in-house precedent: when a console breach turned leaked keys into charges, OpenAI refunded credits drawn in the attack while Anthropic and Google revoked or blocked the exposed keys. OpenAI’s and Anthropic’s own documentation is where the victim-side controls can be checked, and it is narrower than the scale of the risk suggests.
Symantec’s advice line is the practical one, telling organisations to monitor accounts for unusual inference volume or spend, “which is often the first sign that access has been stolen”. The spend dashboard is the detection surface, which makes the spend cap the missing control on the victim side: an agent that runs inside an account with an audit trail and a spend cap is a bounded problem.
OpenAI’s documentation on API key safety tells administrators to set monthly spend limits for the organisation or a project and to enable hard enforcement, and it is blunt about the alternative: “Spend alerts alone don’t stop API traffic.” Hard enforcement is not instantaneous, so spend can slightly exceed the limit, and it can block legitimate requests. If a key is leaked, the documented step is to rotate it immediately and contact support for further investigation. The same documentation describes per-key expiry and an organisation-level maximum key lifetime. It does not publish an automatic refund or credit policy for spending caused by a compromised key, so the practical levers are the cap and fast rotation rather than a dispute route.
Anthropic’s rate-limit documentation sets monthly spend caps by tier: $500 on Start, $1,000 on Build, $200,000 on Scale, and no monthly cap on the custom tier. Reaching the cap pauses API usage until 00:00 UTC on the first day of the next month, and requests return an HTTP 429 error that names the enforced spend limit. Workspaces can set their own limit below the tier cap, and organisation-wide limits always apply. A cap of that kind converts an open-ended theft into a bounded one, and it is documented behaviour rather than an inference.
Logging is the weaker half. In Sysdig’s 2024 work, cloud audit logs recorded model invocation calls but not the prompts or the results, and capturing those required enabling model-invocation logging to a separate store. Attackers looked for accounts where that logging was switched off. The record of what a stolen key actually did may therefore not exist unless someone turned it on in advance.
Two further points belong beside the spend cap. The Cloud Security Alliance note says cloud billing charges may not surface in monitoring dashboards for several hours in default configurations, though real-time budget alerts on AWS, Google Cloud and Azure can shrink that window when they are configured. And providers do sometimes act: Microsoft’s Digital Crimes Unit revoked criminal access, put countermeasures in place and obtained a court order to seize a website in January 2025, against a group that exploited exposed customer credentials to abuse its AI services. That was a response to a different campaign, more than a year before Symantec’s report, and it is the closest thing to a precedent for a provider answering this kind of abuse.
What a defender can check now
Symantec’s recommendations, and the controls providers document, come down to a short list.
Do not expose model servers, AI gateways or MCP endpoints to the internet without authentication. Require authentication on every path, including model-listing and health-check paths, and put services behind a gateway rather than binding a model server to all network interfaces.
Treat AI API keys as high-value credentials. Rotate them, scope them narrowly, keep them out of source code and configuration files, and set a hard monthly spend limit that would be unpleasant rather than ruinous if it were reached. The Aoki infostealer shows how these keys leave: it targets credential and session files used by AI coding assistants, and a poisoned release of a widely used AI gateway library is another route that has already handed attackers live credentials. Neither needs an exploit against a model.
Restrict outbound network access from systems that host AI tools or run tool-enabled agents, since a miner that cannot reach a mining pool is a wasted miner. On cloud workloads, enforce version 2 of the Instance Metadata Service with a hop limit of one. Run tool-enabled agents with least privilege and a human approval step for sensitive actions, and assume an unauthenticated endpoint will be found. Agents that inherit standing access, or that pass instructions to each other without a check, are the same problem seen from the other end, and agents reaching live systems unnoticed are already documented.
The gap in this account is detection. Nobody in it noticed a hijacked model, because there was no hijacked model to notice, only a decoy that recorded what a real one would have been asked to do. Where prompt-level logging is off by default and billing surfaces hours later, the spend cap and the rotation habit do more work than after-the-fact forensics.
