DKnownAI Guard

DKnownAI Guard positions itself at the intersection of AI agent security and prompt injection detection. Unlike traditional content moderation systems that evaluate whether text output is harmful, Guard focuses on a prior question: is someone trying to manipulate the agent itself?

The distinction matters because agentic AI systems increasingly execute code, access files, call APIs, and interact with real infrastructure. A prompt that looks benign in isolation can, in the context of a tool-using agent, trigger destructive actions. DKnownAI Guard's core premise is that the same input can carry fundamentally different risks depending on the agent's capabilities — and that security classification must account for this.

How It Works

DKnownAI Guard exposes a single REST endpoint at open.dknownai.com/v1/guard. Developers send text input and receive one of four classifications: AGENT_HACK (detected manipulation or jailbreak attempt), SYS_FLAG (operational concern requiring scope review), CONTENT_FLAG (harmful content detected), or SAFE (no risk). This four-way taxonomy is the product's central editorial claim: separating deception from content risk allows agents to block hijacking attempts while still processing legitimate requests that happen to contain sensitive keywords.

According to the vendor's published evaluation, the system achieved a 90.4% true negative rate — a metric that speaks to its ability to avoid overblocking ambiguous inputs. A case study documents detection of Grok 4.5 being induced to organize dangerous chemical and biological content into structured documents, demonstrating the gap between surface-level content filtering and manipulation-aware security.

Integration and Scope

The API supports integration through direct HTTP calls or via Guardrail-Bridge, a middleware layer that embeds classification into agent tool chains. This allows developers to insert security checks at decision points — before an agent reads a file, executes a command, or modifies configuration.

DKnownAI Guard's classification is context-aware in the sense that it considers whether a request represents a genuine development task or a disguised attack. A prompt mentioning "API key," for instance, might be a legitimate code review request or a credential theft attempt. The system's design goal is to distinguish these cases, though the available evidence is limited to vendor-authored documentation and self-reported benchmarks.

Limitations

The published evaluation acknowledges that false positives remain an unresolved challenge across guardrail systems, including DKnownAI Guard. The 90.4% true negative rate, while directionally informative, comes from the vendor's own testing and has not been independently verified. Additionally, the API's single-endpoint design means it analyzes text at the input boundary but does not monitor agent behavior during tool execution — a gap that may matter for complex multi-step workflows where attacks unfold across several interactions.

For developers building in the AI Developer Tools space, tools like ExtWise and CodingPlan address adjacent workflow concerns, but DKnownAI Guard occupies a distinct niche: agent-aware security classification that asks not just "is this content harmful?" but "is someone trying to deceive the agent?"