Buying snapshot
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Decision point | What Oracles Technologies states |
|---|---|
Deployment | Three editions: Community runs locally from pip; API routes requests through the hosted Ethicore Engine® platform; Self-Hosted runs the full engine on your own infrastructure. |
Where data goes | Community and Self-Hosted keep data in your environment. The API editions send requests to the vendor's platform, where the threat library and models are managed. |
Threat coverage | Community: seven categories (six OWASP-aligned attack vectors plus an absolute-block child safety category). API and Self-Hosted: 160+ categories. |
License | SDK source under Business Source License 1.1 (changed from MIT), converting to Apache 2.0 on May 7, 2030. The threat library and ONNX models are proprietary and platform-managed, with API access only. |
Installation | pip install ethicore-engine-guardian, with optional extras per provider and for vision, video and voice analysis. Self-Hosted installs ethicore-engine-selfhost. |
Support | Email to support@oraclestechnologies.com; the vendor says you get a direct response from the engineer who built it. |
Standards | The vendor says Guardian SDK is listed in the NIST OLIR catalog, with mappings to NIST Cybersecurity Framework 2.0 and the NIST AI Risk Management Framework 1.0. |
Our take
Guardian SDK is worth a trial if you want one library to screen prompts, model responses and agent tool calls, and you can test its BLOCK and CHALLENGE verdicts against your own traffic. Run that test on the API Free tier, which has the same capability set as the paid API tiers, and pay only once those verdicts hold up on your prompts.
Evaluate Ethicore Engine™ - Guardian SDK
Where it stands out
Several detection layers on each input. The pre-flight gate combines pattern matching (regex with obfuscation normalization), semantic matching with ONNX MiniLM-L6 embeddings for paraphrased attacks, session-level behavioral heuristics for multi-turn escalation and gradual manipulation, and gradient-boosted ML scoring.
Covers the whole agent loop. On the API tier, separate gates check tool calls before they run, scan tool outputs for injected instructions, and validate compiled or parallel execution plans, besides the pre-flight and post-flight checks on model input and output.
Small added latency, by the vendor's figures. Typical latency is ~15ms p99 pre-flight on commodity hardware; OutputAnalyzer and ToolCallValidator each add under 1ms, and ToolOutputScanner adds about 2–5ms depending on output size and format.
Wraps the client you already use. Install extras cover OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral and Perplexity, and an Ollama provider covers local models.
Not tied to Python. Teams on other languages can call the hosted API over HTTP: two endpoints for input and response checks, plus REST endpoints for the agent gates that return an Ed25519 X-Ethicore-Signature header.
LangChain hook-up on the API tier. A GuardianCallbackHandler added to an agent or chain checks prompts, tool calls and tool return values automatically, with an async version for async chains.
Trade-offs to weigh
Agent gates or local data, not both. The tool-call, tool-output and execution-plan gates are API-tier only, so a team that must keep requests on its own infrastructure with Self-Hosted gives up those gates.
The free local edition is a reduced edition. Without an API key the SDK falls back to local hash-based inference with 34 regex patterns, against 1,800+ on the API and Self-Hosted editions.
Source-available, not open source yet. BSL 1.1 allows personal use, internal tools, research, open-source projects and non-competing applications; selling a competing AI security or threat detection product needs the commercial license.
Self-Hosted is a sealed binary. The engine ships compiled, the models ship encrypted, and the license key needs a one-time phone-home activation before an offline grace window applies, so your team cannot inspect or patch the self-hosted engine code.
Pricing and plan conditions
Pro and Team are priced per month. Per the vendor, all API tiers share the full capability set: they differ on request volume, rate limit and support. You choose Free or Pro at registration on portal.oraclestechnologies.com. Self-Hosted is a separate license key, bought and managed on the same portal; it runs unlimited local requests, with a rate limit set by tier from 60 RPM to unlimited.
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Plan | Published price | What it adds |
|---|---|---|
Community | Free | pip SDK the vendor calls open-source (source under BSL 1.1); no API key or account; unlimited local requests |
API Free | $0 | 1,000 requests/mo · 60 RPM |
API Pro | $59.99/mo | 100,000 requests/mo · 600 RPM |
API Team | $299/mo | 500,000 requests/mo · 1,200 RPM · 5 seats · audit log export · priority support · security-questionnaire support pack (pre-filled SIG-Lite / CAIQ answers) |
API Enterprise | Contact us | Custom rate limits + SLA |
Self-Hosted (Bronze / Silver / Gold) | Bronze $29.99/mo · Silver $399/mo · Gold custom | Every detection layer runs in your infrastructure under a license key; agentic pipeline gates excluded (API-tier only) |
Public signals
About these signals
Verified factual signals from public sources. They indicate observable activity or interest, not total adoption, product quality, or cost.
22 GitHub commits 90d123 GitHub stars0 vulnerabilities across 1 package
See all signals from 4 sources
PyPI · ethicore-engine-guardian@2.9.1
Alternatives to Ethicore Engine™ - Guardian SDK
The reviewed substitutes for Ethicore Engine™ - Guardian SDK among the AI security, and what would make each one the better answer.
Other approaches
A different approach to the same problem. Each substitutes only for the workload named beside it.
- EarlyCore
- EarlyCore is chosen instead of Ethicore Engine™ - Guardian SDK for AI-agent programs that require pre-release scanning and production monitoring across Bedrock, Vertex AI, and custom stacks. We recommend EarlyCore over Guardian SDK for organizations that need one security workflow across release evaluation and live agent monitoring rather than a Python-first inline library.
- PromptBrake
- PromptBrake stress-tests an LLM endpoint with 60+ attack prompts across 12 security checks and returns PASS/WARN/FAIL verdicts with evidence, built to run as a CI/CD release gate. Ethicore Engine - Guardian SDK is a pip-installable Python layer that blocks prompt injection, jailbreaks and role hijacking inline, in front of the model call, offline and with no cloud dependency. Both answer AI-application security for the same buyer, so the choice is between repeatable pre-release testing you can export and gate on, and runtime blocking inside the application itself. Recorded as a class-level editorial judgement in ai-security rather than a published head-to-head.Applies to: Defending an LLM application against prompt injection, jailbreaks and unsafe output.
- AgentVault
- Compared with Ethicore Engine™ - Guardian SDK’s request-time threat screening for LLM interactions, AgentVault is the stronger fit when teams need operational control over agent actions and permissions after an agent is deployed. **AgentVault** is a real-time security monitoring and control layer for AI agents, with a focus on OpenClaw and agents such as Claude.