MVP · Live Evidence-first AI security

Security for the teams
shipping AI.

Evidence-first prompt & agent security.

Aegis detects prompt injection, recovers hidden image payloads, and stress-tests your AI against evolving evasions — returning evidence and remediation, not just an unexplained risk score.

<1s
Prompt scan latency
5
Hidden-payload carriers detected
7
Adversarial mutation families
OWASP
LLM Top-10 aligned

Product walkthrough

The security layer
AI teams were missing.

Most security tooling — SIEMs, WAFs, DLP — was built for structured traffic. LLM inference is unstructured, and its failure modes are different: prompt injection, indirect injection through retrieved context, payloads hidden inside images, and jailbreaks that shift faster than static rules.

Aegis was built to close that gap. It analyzes prompts and image carriers, recovers the actual evidence behind a finding, explains why it scored the way it did, and runs adversarial benchmarks so teams can harden defenses before attackers do.

Evidence-first, not black-box Explainable scoring Research-driven

Your entire AI attack surface, in one view.

The Aegis dashboard gives security and ML teams a live view of scan volume, severity distribution, and the highest-risk findings across every prompt, image, and agent run.

Every signal that matters, unified.

A full suite of detection and red-team modules, each built around a real LLM failure mode — and each one returns recoverable evidence, not just a number.

🔎

Prompt Injection Detection

Flags direct and indirect injection across inputs — including multi-turn sequences, retrieved-context injection, and instruction-override patterns.

DirectIndirectMulti-turn
Shipped
🧠

NLP Threat Analysis

Semantic NLP engine that reads intent — catching jailbreaks, policy-violating requests, and obfuscated instructions that simple keyword or regex filters miss.

SemanticIntentJailbreak
Shipped
🧩

Hidden Payload & Steganography

Recovers payloads from PNG metadata, appended/trailing bytes, and PNG/BMP LSB channels, with chi-square anomaly checks — then re-scans recovered text.

MetadataLSBOCRChi-square
Shipped
⚔️

Red Team Mode

Generates evolving adversarial prompts to probe your model, surfaces successful evasions, and exports them as regression cases so defenses keep improving.

Adversarial probesEvasion finderRegression export
Shipped
🎯

Adversarial RL Engine

The reinforcement-learning core behind red-team mode — 7 mutation families (role override, obfuscation, multilingual, emoji/unicode, tool-use coercion, exfil framing) with a full per-step trace.

7 familiesStep traceBefore/after score
Shipped
🔒

Explainable Scoring

Surfaces score drivers, evidence-vs-suspicion distinction, and OWASP LLM references so every finding carries context and remediation.

Score driversOWASP refsRemediation
Shipped
📋

Scan History & Replay

Tenant-scoped scan history with severity, mode, and time filters, a detail drawer with reasoning, and one-click replay — IDOR-guarded by design.

FiltersReplayTenant-scoped
Shipped
🤖

Agent Orchestration (ADK)

Env-gated Google ADK orchestrator exposing typed tools. Reports are grounded only in tool outputs and cite their evidence — no invented findings.

Typed toolsGroundedCited
Preview

Everything a security team needs, in one engine.

Aegis is more than a scanner — it's a workflow. From first scan to a tracked, replayable security posture.

01

Command Center

A live dashboard of scan volume, severity distribution, and your highest-risk findings across prompts, images, and agent runs.

02

Prompt Scanner

Submit a prompt or conversation and get an explained verdict — risk score, severity, reason codes, and remediation in under a second.

03

Image & Stego Lab

Inspect image carriers for hidden instructions — metadata, trailing bytes, LSB channels, and chi-square anomalies, with recovered text re-scanned.

04

Red Team Runner

Launch adversarial RL runs against your model, watch the per-step mutation trace, and see exactly where defenses hold or break.

05

Scan History

Every scan is logged, filterable by severity, mode, and time, with a detail drawer for reasoning and one-click replay — tenant-scoped.

06

Threat Memory

Successful evasions are saved as regression cases, so a vulnerability you find once is something you test for every time after.

07

Reports & Export

Generate evidence-backed reports and export findings — including regression packs — for audits, stakeholders, and CI pipelines.

08

Agent Orchestration · preview

An env-gated ADK orchestrator that calls Aegis tools and produces grounded, cited reports — never inventing findings beyond tool outputs.

From input to evidence-backed report.

01

Submit a prompt, image, or agent run

Send input through the dashboard or the REST API. Text prompts, image carriers, and multi-turn conversations are all supported.

02

Detection runs in parallel

Aegis runs injection detection, hidden-payload recovery, statistical anomaly checks, and adversarial mutation — recovering the actual evidence behind each signal.

03

Act on the explained verdict

Get a 0–100 risk score, severity, recovered evidence, reason codes, and remediation — with full history and replay for every scan.

scan-response.json
{
  "scanId": "SCN-7F32-AE",
  "detected": true,
  "riskScore": 84,
  "severity": "CRITICAL",
  "carrierMethod": "png_metadata",
  "evidenceType": "recovered_text",
  "recoveredText": "ignore previous… exfiltrate keys",
  "owaspRef": "LLM01: Prompt Injection",
  "remediation": "strip metadata, sandbox tool calls",
  "latencyMs": 612
}

Start free. Scale when you ship.

Aegis is in early access. Request an invite to start on a genuinely free plan — no card required — and talk to us for team and enterprise needs.

Free
$0/early access
A genuine free tier during early access — scan real prompts, no card required.
  • Text prompt injection scanner
  • Basic remediation guidance
  • 250 scans / month
Request access
Team
Let's talk
For security & ML orgs that need shared workspaces and full coverage.
  • Everything in Pro
  • Multi-user workspace & roles
  • Threat Memory & scheduled scans
  • Report export & API keys · 100,000 scans/mo
Contact us
Enterprise
Custom
For regulated orgs needing private deployment, SSO, and data residency.
  • Everything in Team
  • SSO / SAML / OIDC & RBAC
  • Private GCP / VPC deployment
  • Audit logs, tenant isolation & custom SLA
Contact sales

AI Prompt Injection & Hidden Payload Assessment

A hands-on engagement: we red-team your AI application against injection, steganographic payloads, and agent-tool abuse, then deliver an evidence-backed report with remediation.

Custom scope
Talk to us

Compare plans

Feature Freeearly access Pro$49/mo Teamcontact Enterprisecustom
Detection
Prompt injection scanner
Remediation guidance
Detailed score reasoning
Unicode / emoji / multilingual (NLP)
Saved scan history & replay
Image & steganography scans
Red team & testing
Red team / RL adversarial runs
Threat Memory
Scheduled scans
Report export
Workflow & limits
Monthly scans included 250 10,000 100,000 Custom
Multi-user workspace
Multiple API keys
Security & deployment
SSO / SAML / OIDC & RBAC
Audit logs & tenant isolation
Private GCP / VPC deployment
SLA & dedicated support Community Email Priority Custom SLA

Evidence over marketing claims.

We don't ship testimonials we can't stand behind. We ship detectors, traces, and reports you can verify yourself.

8

Modules shipped

Injection, NLP analysis, steganography, red-team, RL engine, scoring, history, and ADK orchestration.

5

Hidden-payload carriers

Metadata, appended bytes, PNG & BMP LSB, plus chi-square anomaly detection.

7

Mutation families

Role override, obfuscation, multilingual, emoji/unicode, sandwiching, tool-use coercion, exfil framing.

OWASP

LLM Top-10 aligned

Findings map to recognized LLM risk categories with cited references.

Ready to secure what you ship?

If you're deploying AI applications and want clear evidence of how your defenses behave under real pressure, Aegis is a good place to start.

Let's talk AI security.

Whether you're exploring Aegis, have a technical question, or want to scope an assessment — reach us directly. We read every message.