CalypsoAI’s July 2025 release introduces new Red-Team attack capabilities, key usability improvements, and expanded campaign insights, reinforcing our commitment to securing AI with real-time, actionable protection.

Inference Red-Team

Style Injection Joins the Arsenal

This month’s Inference Red-Team release includes a new attack pack with 10,500+ new malicious prompts and introduces a new vector: style injection.

This jailbreak method manipulates formatting and writing style (e.g., “use no punctuation,” “five-word sentences only”) to subtly coax models into bypassing refusals. It’s another example of how attackers are getting creative and why dynamic, evolving defenses are essential.

We’ve also renamed our monthly “prompt packs” to attack packs, which is a small but important step to align terminology with our offensive security approach.

Improved Usability in Raw Data Table

We’re continuing to improve usability in the raw data table for Red-Team reports based on customer feedback. In this release you’ll see:

• Better filtering and sorting capabilities
• Visual indication of the number of results in the table (before and after filtering)
• Automatic loading of more results as you scroll

More Visible Campaigns

We’ve made it easier to track and manage Red-Team campaigns. You’ll now see a dedicated campaign column on the Reports page, along with clickable campaign names in both the report UI and the downloadable file. This allows you to quickly jump to the exact campaign you’re reviewing. We’ve also added new columns in the raw data view for report name and campaign ID, giving you more context at a glance.

Error messaging for reports

We’ve started improving error messages in Red-Team reports to make troubleshooting easier. In this release, we’ve added a new “schedule error” category to help identify issues during scheduled campaign runs. This error appears when there’s a problem connecting to one of your target models or applications. If you see it, simply test the connection and resolve any issues. Red-Team will then automatically pick back up with the next scheduled run once the connection is restored.

Inference Defend

Custom Scanner Details

July’s Inference Defend release expands what you can do with custom scanners. As a first step, you can now click on any scanner name in the table to open a detailed view, which shows the scanner’s tags and full description. This provides more visibility into how each scanner works and what it’s designed to catch.