Capabilities ApexAI Parley Guard
Capability · AI security
Stop instructions that were never yours.
An AI model follows the words in front of it, whoever wrote them. A user can try to talk it out of its rules, and a web page, document or tool result can carry hidden instructions of its own. ApexAI Parley Guard reads what goes into a model and flags text that is trying to take control of it.
Status: Preview Parley Guard runs in observe mode: it scores and explains prompts and tool results but does not block; inline enforcement arrives with ApexAI Causeway.
Why prompt injection is different
Prompt injection is not a bug in one product. It follows from how language models work: instructions and data arrive as the same kind of text, so a model cannot always tell which is which. A direct attack comes from the person typing, asking the model to ignore its rules or to play a role that has none. An indirect attack is quieter: the instructions sit inside content the model reads on someone else’s behalf, such as a search result, an e-mail, a document or the output of a tool.
Indirect injection matters more every month, because AI agents now read outside content and then act: they call tools, query systems and send messages. A single poisoned tool result can turn an assistant into an insider. Security teams need a check that sits in the path of that content and does not slow it down.
What Parley Guard checks
Direct prompts
Prompts sent to a model are checked for injection and jailbreak attempts, such as requests to drop the system instructions or bypass safety rules.
Tool and MCP results
Content returned by tools, MCP servers and retrieval is checked for instructions aimed at the model, the main path for indirect injection.
A reason with every flag
Each result comes with a score and the reasons behind it, so a reviewer can see why the text was flagged.
Talking about attacks is fine
An analyst asking how an attack technique works is asking about an attack, not making one. Parley Guard is tested so that security questions are not flagged.
Local and fast
The guard runs locally, close to the traffic, and is light enough to sit in the request path without a noticeable delay.
ApexAI protects itself too
Alerts and logs can carry attacker-written text. ApexAI is designed to place the same guard in front of its own AI, so that evidence cannot instruct the analyst.
From observe to block
Like every ApexAI control, Parley Guard is switched on in stages, and each stage is a decision you make.
01
Observe
Prompts and tool results are scored and recorded. Nothing is changed or stopped, so you can see what would have been flagged.
02
Sanitise
Flagged content is cleaned or held back from the model, while the request continues.
03
Block
Requests that clearly carry an injection are refused, with the reason recorded in the audit trail.
What is in place and what is next
| Part | What it does | Status |
|---|---|---|
| Injection and jailbreak check on prompts | Flags direct attacks on the model, with reasons | Preview, observe mode |
| Injection check on tool and MCP results | Flags instructions hidden in content the model reads | Preview, observe mode |
| Inline enforcement on the gateway | Sanitise or block in the request path | In development with ApexAI Causeway |
| Sensitive data and output checks | Keeps personal data and secrets out of prompts and answers; checks what the model returns | Roadmap |
Questions
What is the difference between direct and indirect prompt injection?
Direct injection comes from the person writing the prompt. Indirect injection is hidden in content the model reads, such as a web page, a document or a tool result, and is aimed at whoever’s AI reads it next.
Does Parley Guard send our prompts to an outside service?
No. It runs locally, next to your traffic. On the gateway, prompt and response content is not stored by default; only what is needed for audit is kept.
Will it block my security team’s questions about attacks?
It should not, and it is tested for exactly that: questions about attack techniques are normal SOC work. In observe mode you can check this on your own traffic before enabling anything.
Can I use it without ApexAI Causeway?
Today it runs in observe mode inside the ApexAI platform. Inline sanitising and blocking are planned through Causeway, the gateway that AI and MCP traffic passes through.
See ApexAI on your own alerts.
A walkthrough with your team, in observe mode first: ApexAI scores and proposes, nothing acts on its own.