Search K
Appearance
July 28, 2026 - SaaS - v10.81.3
INFO
AI Guardrails now supports scanning responses.
This release includes updates in the following areas for F5 AI Security:
AI Guardrails
Scanning responses
AI Guardrails scans prompts by default. You can now also scan responses. Response scanning prevents restricted information from reaching users.
This release introduces two-way scanning for default guardrails, which enables deeper inspection of inbound and outbound AI traffic. This feature may introduce additional latency, which typically ranges from 10–30 seconds (average ~20 seconds). Third-party AI models account for most of the latency. F5's processing adds approximately 500 milliseconds. You can disable two-way scanning if latency becomes a concern.
AI Red Team
Attack pack
The August 2026 attack pack includes a Morality Dilemma, which manipulates the model's ethical alignment by embedding misleading or competing moral norms into the prompt.
Echo chamber attack vector
The echo chamber attack plants 'poisonous seeds' (innocuous sentences that contain keywords related to a harmful objective). Over multiple turns, it manipulates the model into developing those ideas, gradually steering the conversation toward harmful content. To reduce risk, deploy guardrails that monitor keyword accumulation and semantic drift across multi-turn conversations. Apply safety filters to user prompts and model responses at every turn.
Improved runtime
Process optimizations reduce total runtime for red-team campaigns by up to 10%.
Resolved issues
This release fixes the following issues:
- Raw data view for operational attacks now includes full details.
- Custom guardrail message now shows number of affected projects.
- Outcome analysis now highlights correct text in JSON response.
- Audit log filters now update the URL without a full page reload.
Doc updates
Since the last SaaS release on July 8, AI Security has added the following documents:
INFO
The table of contents has been extensively changed. Content is now organized by function, such as Get started with AI Security, AI Guardrails, AI Red Team, and About the API.