RXed AI News

We x-ray the AI industry.
@RXed_EU
Gr Compositions

Guardrails

Live checks on every input and output, plus a human gate before risky actions.

Compositions · ValidationScored on 62/62 audited toolsMedian 6/10Best Gong 9

Guardrails sit between your AI and the outside world and run on every single request: they hide sensitive data going in, filter harmful or leaky answers coming out, catch hidden instructions planted in the text your agent reads, and pause the agent for a human before anything irreversible. Without them an agent will repeat a customer's card number in an email summary, or follow an instruction buried in a support ticket. From 2 August 2026 human oversight of high-risk AI is not a preference in the EU, it is EU AI Act Article 14, with fines up to 3% of global turnover.

You need this when

  • Your AI reads text you did not write: inbound email, support tickets, uploaded files, web pages.
  • It can act outward: send mail, move money, call an API, change a customer record.
  • You touch personal or regulated data, or you sell into the EU where Article 14 oversight applies.

You can skip it when

  • Solo drafting and brainstorming with no customer data: the model's built-in safety filters are enough.
  • Read-only Q&A over your own public docs: spend the effort on retrieval quality, not policy layers.
  • A one-week prototype: run shadow mode (AI drafts, human sends) instead of buying a guardrails platform.

Flashcards

Check yourself

1. When do guardrails do their work?

All four types run at runtime, on every request. Testing tells you if they work; it does not substitute for them.

2. A support ticket contains hidden text telling your agent to email the client list. This is:

Prompt injection hides instructions in content the model reads. It beats unprotected systems over 50% of the time.

3. In the 62-tool audit, which pair scores lowest on Guardrails at 2?

Grok (xAI) and DeepSeek both score 2. DeepSeek has no moderation API and stores hosted prompts in China.

4. You want guardrails without a per-call vendor bill. Best move?

Open-source rails run on your own compute and you control the checks. 2FA and prompt wording are not runtime content controls.

5. From 2 August 2026, EU AI Act Article 14 requires what for high-risk AI?

Article 14 turns human oversight into a legal duty, with fines up to 15 million euro or 3% of global turnover.
5 questions · nothing is tracked, nothing is sent anywhere.

Cheat sheet

  • Four types: input filters, output filters, injection defense, human approval gates.
  • Guardrails run live on every request, not just during testing or evaluation.
  • Prompt injection succeeds over 50% of the time against systems with no protection.
  • EU AI Act Article 14: human oversight of high-risk AI is law from 2 August 2026.
  • Free or self-hosted: LLM Guard, NeMo Guardrails, Presidio, n8n's Guardrails node.
  • Context-layer rails filter before the model sees data; model filters only run after.

Who actually does this well

Best on this elementScoreWhy it scored that
Gong9Agents inherit existing security permissions, run with scoped access, configurable approval gates and human oversight by design. Add ISO/IEC 42001 for AI management, SOC 2 Type 2,
Gemini / Google8.5Model Armor screens prompt injection, PII leaks and toxic content in real time, on top of configurable safety filters, ADK deterministic guardrails and human-in-the-loop pauses.
Microsoft 365 Copilot8.5Best-in-class enterprise controls: Purview DLP on prompts, sensitivity labels, watermarking for sensitive material, admin-governed environments and approval steps in workflows.
Claude / Anthropic8Fable 5 ships dedicated safety classifiers, streaming-refusal handling is documented, and Managed Agents adds permission policies plus credential vaults — integrators must handle r
Cognism8Compliance is the product: GDPR/CCPA legitimate-interest processing with documented assessments, Do-Not-Call scrubbing across 15+ territories, contact notification within GDPR time
Glean8A dedicated 'Protect' safety line — sensitive-content detection, prompt-injection detection, human-in-the-loop write confirmation — though a May 2026 independent audit found some P
Sierra8Deterministic guardrails plus layered "supervisor" agents for input filtering and output interception are genuinely sophisticated — but the December 2025 Gap.com misconfiguration s
Cursor7.5Cursor 3's preview-then-approve diff gates, scriptable Hooks, sandboxed terminal, model block lists and org-wide privacy mode are real runtime controls; compliance surface is SOC 2

And the other end of the same column:

WeakestScoreWhy it scored that
DeepSeek2No moderation API or policy controls, and the hosted service stores prompts in China per its own privacy policy, with South Korea's PIPC documenting undisclosed transfers — self-ho
Grok (xAI)2The January 2026 record is disqualifying without an external safety layer: Grok Imagine produced sexualized imagery of minors and nonconsensual deepfakes at scale, restrictions lan
Higgsfield2NSFW filtering exists but is inconsistent (legitimate fashion/lingerie brands report false flags) while Forbes documented the platform distributing non-consensual deepfakes of Sydn
Luma AI2No customer-exposed content filters, PII detection, or policy enforcement surfaces documented; a gap for a generative platform handling creative content.

Scored on 62 of 62 audited tools. Every score links to the full audit and its reasoning.