Prompt adware: how to audit AI agents for hidden instructions
Key takeaways
- Prompt adware changes an agent's trusted natural-language instructions, not its code
- Haiou injected an ad for a third-party AI gateway via global Codex instructions
- Teams should inventory agents, Skills and config files, and review them for persistent changes
- CPR saw no data theft, but the same persistence could be abused maliciously
After reading this guide, an IT team will be able to inventory the AI agents in their organisation, find the prompts, Skills and configuration files that shape their behaviour, and spot the kind of persistent instruction tampering Check Point Research (CPR) has named "prompt adware".
What is prompt adware?
Prompt adware is CPR's term for modifications that persistently change an AI agent's trusted natural-language instructions, such as global prompts, Skills or configuration files, for commercial or behavioural purposes. Traditional adware uses software to show advertising. This version changes the instructions that help decide how the agent behaves, with no executable payload to catch.
The example is Haiou, an OpenAI Codex jailbreak bundle. It uses several prompt-injection techniques to bypass Codex's restrictions. CPR found it also rewrites Codex's global agent instructions so that ordinary greetings trigger an advert for a third-party AI gateway. The gateway is not known to be malicious. The mechanism is the problem.
What you need before starting
- Difficulty: moderate. No exploit skills required, but you need admin access to developer machines or endpoint management.
- Access: authority to read agent configuration files and instruction files on managed devices, plus a list of teams using coding agents.
- People: someone from security, someone from engineering, and ideally whoever owns procurement of AI tools.
- Cost: none for the audit itself beyond staff time.
Step-by-step process
Step 1: Inventory your AI agents
CPR's first recommendation is to understand which AI agents and third-party components are in use across the organisation. Work out which agents and third-party components are in use across the organisation. Record the agent, the version, who uses it and what it can reach: source code, internal documents, cloud credentials.
Step 2: Locate the instruction layer (1 to 2 hours)
For each agent, find where trusted natural-language instructions live. Haiou targeted Codex's global agent instructions, so start with global and per-project instruction files, any installed Skills, and configuration files. The aim is a list of every file that can change what an agent does without anyone touching executable code.
Step 3: Review what is in them (1 to 3 hours)
Read the files. Look for text that does not belong: promotional wording, references to external services or gateways, instructions triggered by ordinary phrases such as greetings, and anything telling the agent to recommend, favour or route work to a named product. Haiou's tell was exactly this kind of trigger-word behaviour. The ad was simple, but CPR notes the same persistence could steer recommendations or bias technical advice less visibly.
Step 4: Check where each component came from
For every third-party Skill, bundle or gateway, record who supplied it and why it is there.CPR also points to separate academic findings about other LLM gateways that intercepted user data, exposed cloud access keys and sometimes injected tool calls. Gateways can handle your data, so weigh that risk when approving one.
Step 5: Set an approved list
Decide which AI tools, Skills, gateways and external components are approved for enterprise use, and publish it. Anything outside the list gets reviewed before use. This is the governance step CPR recommends around third-party AI tools.
Step 6: Monitor for persistent changes
Put the instruction and configuration files under change monitoring. A global instruction that quietly gains a new line is the Haiou pattern. The aim is to account for persistent changes to the instructions and configurations that shape agent behaviour.
Step 7: Review access alongside risk
CPR advises weighing what an agent can access against what can influence its actions. Where an agent touches proprietary code or credentials, tighten its permissions. CPR did not see Haiou stealing corporate information, but confidential data exfiltration is the business risk it flags if a similar technique were used maliciously.
Common mistakes to avoid
- Checking only the model. CPR's point is that prompts, Skills and configurations are part of the trust chain too. A vetted model with an unvetted instruction file is still exposed.
- Searching only for malware. Haiou's advertising behaviour came through natural language and configuration, not a traditional executable payload that security tooling may more readily recognise.
- Assuming a benign gateway means a safe pattern. The gateway Haiou promoted is not known to be malicious. The persistent edit to trusted instructions is still the risk.
- Relying on the agent to defend itself. Claude Code flagged and refused the injected instruction in CPR's testing, but that is one agent in one test, not a control.
Tips that make a difference
- Read the jailbreak-style bundles your developers actually download, not just the official tooling. Haiou was found by examining what the bundle changed in the agent, rather than only asking whether its jailbreak worked.
- Test with ordinary phrases. If a greeting produces unexpected promotional content, check the instruction files for planted text.
- Keep a baseline copy of each instruction file so diffs are quick.
The UK picture
Next steps
Fold agent instruction files into existing change management. Extend third-party risk reviews to cover Skills and gateways. For related reading on instructions hiding where defenders do not look, see our report on hidden commands found in website headers targeting AI, and for the downside of broad agent access, OpenAI's agent breached an Australian health portal.
Haiou's advert is harmless in itself. What matters is that a third-party bundle persistently altered instructions an agent trusts, and that few teams would see it. CPR's guidance is to gain visibility, review the components around agents, govern third-party tools, monitor instruction changes and weigh access against influence.
Summary
You now have a repeatable audit: inventory agents, locate and read their instruction layers, vet third-party Skills and gateways, set an approved list, monitor for persistent edits and tighten access. That covers the prompt adware technique CPR documented and the more harmful variants it warns could follow.
Frequently asked questions
What is prompt adware?
Prompt adware is Check Point Research's term for persistent changes to an AI agent's trusted natural-language instructions, such as global prompts, Skills or configuration files, made for commercial or behavioural purposes rather than through executable code.
What is Haiou?
Haiou is an OpenAI Codex jailbreak bundle that also edits Codex's global agent instructions so common greetings trigger an advert for a third-party AI gateway. Check Point did not find the gateway to be malicious.
Did Haiou steal data?
No. Check Point Research did not observe Haiou stealing corporate information. The concern is what the same persistence mechanism could enable if trusted instructions or components were deliberately abused.
Does prompt adware affect Claude Code?
In Check Point's testing, Claude Code identified the attempted prompt injection and declined to follow the injected instruction. The developer behind Haiou had experimented with Claude Code.