Verosek Shield scanning AI agent requests and responses for prompt injection and unsafe tool outputVerosek Shield

24 checks. One layer. Every request scanned.

Input. Output. Tool output. Session drift. Deterministic where possible. Offline ML everywhere.

Full Check Catalog

Every check. Every scan point. No black box.

CHK-013S1
Prompt injection in user inputOffline multilingual classifier
PRE-LLM
CHK-014S1
Jailbreak & role-play attemptsHeuristic + classifier ensemble
PRE-LLM
CHK-015S1
PII in user inputNamed-entity recognizer
PRE-LLM
CHK-031S1
PII in model outputNamed-entity recognizer
POST-LLM
CHK-032S1
Toxic or unsafe outputOffline safety classifier
POST-LLM
CHK-021S2
PII in tool outputNamed-entity recognizer
POST-TOOL
CHK-022S2
Secrets in tool outputEntropy + pattern match
POST-TOOL
CHK-024S2
Off-topic tool callPer-desk topic centroids
PRE-TOOL
CHK-025S2
Unapproved tool invocationPolicy engine
PRE-TOOL
CHK-041S3
Session drift over turnsRolling embedding monitor
POST-LLM
CHK-042S3
Data-exfiltration patternSequence anomaly model
POST-TOOL
CHK-043S3
Cross-tenant leakageTenant-scoped ACL check
PRE-LLM
Request Lifecycle

1 Request – 4 scan points

Most guardrails scan the prompt and stop. Shield scans every point where untrusted text can enter — including the ones after the model has already replied.

Before the model sees anything

PRE-LLM

Prompt injection, jailbreak, PII, secrets, scope creep — CHK-013, 014, 015, 016, 024

When your MCP server replies

POST-TOOL

Indirect injection hiding in a SQL row or Slack message — CHK-020, 021

Before the user sees the answer

POST-LLM

Toxicity, leaked PII, leaked secrets, ungrounded claims — CHK-017, 018, 019, 023

Across every turn

SESSION

Cumulative exfiltration drift the single-turn scans can't see — CHK-022

Every signal is local.
The gateway never imports torch.

No prompt, no PII, no tool output ever leaves your network to get scanned. All four systems below run on infrastructure you control.

Offline Classifiers

Prompt injection, jailbreak, toxicity — runs on-prem, no outbound.

Multilingual PII Engine

6 languages, 4 redaction modes (mask / redact / tag / block).

Session Drift Counters

Cumulative PII, URL, and byte-count tracking across turns.

Grounding Verdict

Local, async, off the hot path. Never delays the response.

Frequently Asked Questions

Filters watch the prompt and the reply. Real agent attacks arrive inside tool results and build up across a session. Verosek watches those boundaries too, which is exactly where MCP attacks land and where most products are blind.

You prove safety before you enforce. A check can run silently against live traffic first, then switch on only once it stops flagging real users. Protection without the false-positive tax.

No. Scanning happens inside your own environment with no outbound calls. Your prompts and data never leave, even fully air-gapped.

Detection is layered, runs where attackers cannot see or probe it, and resolves deterministically rather than by a model they could coax. We keep the exact checks private on purpose. You cannot defeat what you cannot inspect.

Two lines to migrate. Every SDK to every provider.