F5 AI Guardrails (formerly CalypsoAI) Review

Industrial cybersecurity

TL;DR

Last updated: November 3, 2026. Testing track: Documentation Reviewed.

F5 turned CalypsoAI into F5 AI Guardrails and F5 AI Red Team, made them generally available in January 2026, and wrapped them in an F5 AI Security Platform with SurePath AI’s network-based discovery. It is strongest for enterprises protecting their own AI apps and agents with self-hosted or air-gapped options. The newer agentless Workforce AI Security product extends coverage to employee AI use, but it is recent and lightly documented.

Ownership: CalypsoAI acquired by F5: acquisition announced September 2025; closed 2025-09-26 for $145.2M in cash per F5 Form 10-Q (quarter ended 2026-06-30). F5 also closed the acquisition of SurePath AI on 2026-06-15 for $50.1M (same 10-Q), announced 2026-06-22.

Score: 6.8 / 10.

Initial review November 3, 2026.

How This Review Was Conducted

This review is based on F5 AI Guardrails and AI Red Team’s public documentation, product pages, pricing pages and press coverage as of November 2026. We have not had lab access or a vendor demo.

Scores use the same published rubric and formula as every review on this site. Sources are listed at the end of the review.

Review methodology and evaluation process

Score breakdown

Dimension

Coverage breadth

Detection accuracy

Deployment friction

Policy & control depth

Framework alignment

Pricing transparency

Support & documentation

Weight

20%

20%

15%

15%

10%

10%

10%

Score

8

7

6

8

5

5

7

Notes

Custom apps, agents, MCP and now agentless workforce coverage, with the workforce piece very new.

Large attack library and an efficacy claim, but the independent tester is unnamed.

Flexible SaaS or self-hosted, but self-hosting needs Kubernetes and workforce needs traffic mirroring or redirects.

Custom and plain-language guardrails, block or flag, redaction, group-based and tool-call controls; no coaching.

Regulatory templates and EU AI Act references, no documented OWASP or NIST mappings.

Marketplace placeholder figure only; real pricing by private offer.

Solid docs and release notes, community playbooks, base support included.

Coverage breadth

Weight
20%

Score

8

Notes

Custom apps, agents, MCP and now agentless workforce coverage, with the workforce piece very new.

Detection accuracy

Weight
20%

Score

7

Notes

Large attack library and an efficacy claim, but the independent tester is unnamed.

Deployment friction

Weight
15%

Score

6

Notes

Flexible SaaS or self-hosted, but self-hosting needs Kubernetes and workforce needs traffic mirroring or redirects.

Policy & control depth

Weight
15%

Score

8

Notes

Custom and plain-language guardrails, block or flag, redaction, group-based and tool-call controls; no coaching.

Framework alignment

Weight
10%

Score

5

Notes

Regulatory templates and EU AI Act references, no documented OWASP or NIST mappings.

Pricing transparency

Weight
10%

Score

5

Notes

Marketplace placeholder figure only; real pricing by private offer.

Support & documentation

Weight
10%

Score

7

Notes

Solid docs and release notes, community playbooks, base support included.

What it does well

Red team findings convert into active guardrails.

Red teaming and runtime guardrails are tightly linked, with red team findings convertible into active guardrails (F5 blog, 2025-09-29).

Supports self-hosted and air-gapped deployment.

Self-hosted, on-premises and air-gapped deployment for regulated and sovereign environments (F5 press release 2026-06-22; AWS Marketplace listing).

Attack library has 140,000+ patterns.

Large and growing attack library of 140,000+ patterns (F5 press release 2026-06-22).

Agentless workforce coverage via network traffic.

Agentless workforce coverage via network traffic avoids deploying browser extensions or endpoint clients (F5 Workforce AI Security page). Backed by F5's application delivery footprint and financial resources (F5 10-Q).

Where it falls short

Efficacy claim lacks named methodology.

The 98.2% efficacy claim does not name the tester or methodology (F5 press release 2026-06-22).

Framework mappings not documented in primary sources.

Framework mappings (OWASP, NIST AI RMF, MITRE ATLAS) are not documented in F5 primary sources.

Workforce AI Security has limited documentation.

Workforce AI Security is new (GA stated as October 2026) and has limited public documentation. SIEM integration is not documented. Marketplace pricing signals a high entry point ($100,000 per 12 months placeholder), with real pricing via private offer.

Best fit

Large or regulated enterprises building their own AI applications and agents that need model-agnostic runtime guardrails and automated red teaming they can run self-hosted or air-gapped, particularly existing F5 customers who want AI controls alongside their application delivery and security stack.

Poor fit

Small and mid-sized organizations with modest budgets, or teams mainly needing browser-level employee AI DLP with coaching, M365 or Google Workspace app governance, or documented framework mappings. Alternatives: Harmonic, Nudge or Portal26 for workforce AI usage control; Cisco AI Defense for documented OWASP and NIST mappings; Check Point (Lakera) for runtime guardrails with published detection metrics.

Pricing transparency

Not published on the F5 site. AWS Marketplace listing for F5 AI Guardrails (EKS container) shows a contract of $100,000 per 12 months, described as a placeholder until F5 configures a private offer; base support included, consulting and TAM extra.

Alternatives

See how it compares in our Best LLM Security Tools 2026 ranking. The poor-fit section above names alternatives for buyers it does not suit, and our methodology explains how every product is scored.

What We Would Test in the Lab

If F5 AI Guardrails and AI Red Team grants lab access, we would run the following scenarios and upgrade this review to Lab Tested.

Prompt injection and jailbreaks.

Direct and indirect prompt-injection scenarios from public OWASP and Lakera-style test sets against a representative LLM app.

Sensitive data leakage.

The standard sensitive-data set in prompts and model responses.

Policy enforcement.

Block, warn, redact and allow behaviors against a configured policy.

Audit logging.

What is logged, what is not, and retention behavior.

SSO integration.

Microsoft Entra ID and Okta where supported.

Deployment.

Time from access to enforced policy in a representative environment.

Latency.

Added latency on standard prompt sizes.

Feature coverage at a glance

Based on public documentation. “Not documented” means we could not find it in public material, not that the product lacks it.

CapabilityF5 AI Guardrails and AI Red Team
Inline prompt DLPYes. Guardrails block data leakage; Workforce AI Security redacts PII, financial data and IP in prompts.
Shadow AI discoveryYes. Network-based discovery of sanctioned and shadow AI, private models and MCP connections (SurePath).
Blocking unapproved AI toolsYes. Workforce AI Security blocks high-risk or prohibited AI services and routes to approved models.
File upload inspectionPartial. Workforce press release lists controls by file upload; content inspection of files not described.
Secure AI workspace (sanctioned chat)Not documented. Routing to approved enterprise models is described; no vendor chat interface documented.
Built-in AI vendor risk scoringNot documented. Usage and intent classification described, not AI vendor risk ratings.
Microsoft 365 connected AI app review and controlNot documented. Not described.
Google Workspace connected AI app review and controlNot documented. Not described.
Governance scorecard, attestation and board reportingPartial. AI Governance pillar, audit trails and regulatory templates; no scorecard or board report documented.
AI agent and MCP securityYes. Traces MCP tool calls and agent sessions; Workforce can allow, block or modify tool calls.
Prompt injection defenseYes. Runtime guardrails block prompt injection and jailbreaks.
AI red teamingYes. F5 AI Red Team with agentic attack campaigns; findings convert into guardrails.
SIEM integrationNot documented. No SIEM integration found in F5 sources reviewed; Splunk AI Guardrails Add-on does not list F5.

Recent changes

  • 2025-09-26: F5 closed CalypsoAI acquisition ($145.2M cash, F5 10-Q).
  • 2025-09-29: F5 introduced F5 AI Guardrails and F5 AI Red Team.
  • 2026-01: F5 AI Guardrails and AI Red Team generally available as part of F5 ADSP.
  • 2026-03-30: SaaS release added Gemini and Azure OpenAI connections and flagOnly blocking parameter.
  • 2026-06-15: F5 closed SurePath AI acquisition ($50.1M, F5 10-Q); 2026-06-22 F5 AI Security Platform launched.
  • 2026-09-09: F5 Workforce AI Security announced, GA stated as October 2026.

Sources

Deployment and coverage

AI Guardrails: SaaS API, or self-hosted container on Kubernetes or OpenShift via the F5 AI Security Operator (Amazon EKS listing); cloud, hybrid, on-premises and air-gapped supported. Workforce AI Security and AI Discovery: agentless, passive out-of-band analysis of mirrored network traffic or network redirects, integrating with existing SASE and identity providers (Azure AD, Okta); no endpoint client. Typical rollout time not documented.

Custom LLM apps, models and agents (model-agnostic, connections to providers including Gemini and Azure OpenAI); MCP tool calls and agent sessions; network-level discovery of sanctioned and shadow AI apps, private models and MCP connections (SurePath); employee AI use across browsers, CLIs, coding agents, MCP clients and self-built tools via Workforce AI Security; MuleSoft Agent Fabric integration (per press coverage).

FAQ

What happened to CalypsoAI?

F5 closed its acquisition of CalypsoAI on 2025-09-26 for $145.2M in cash and rebranded its products as F5 AI Guardrails and F5 AI Red Team, generally available from January 2026.

Yes. F5 documents cloud, hybrid, on-premises and air-gapped deployment, including a Kubernetes container on Amazon EKS and an OpenShift operator.

F5 does not publish pricing. The AWS Marketplace listing shows a $100,000 per 12 months placeholder, with actual pricing set by private offer.

F5 Workforce AI Security, announced 2026-09-09, discovers and controls employee AI use agentlessly from network traffic, including redaction and blocking.