We’re excited to introduce our latest updates, packed with major platform enhancements to our red team capabilities, AI security scoring, attack arsenal, and scanner customization. These releases also include new scheduling features, expanded scanner controls, updated prompt injection defenses, and a fresh rebrand for CalypsoAI.
Innovation in AI moves fast, and we need security to move just as quickly. That’s why we release new features every two weeks, ensuring our customers stay ahead of evolving threats and industry advancements. Here’s what’s new.
Inference Red-Team Enhancements
Red Team Reports – Actionable Insights at Your Fingertips
When red-teaming AI models, the most valuable output is the report. We’ve streamlined this process with our new Red Team Reports, delivering:
- A concise summary of campaign results with per-prompt details and downloadable spreadsheets.
- The ability to attack multiple models at once, with both aggregated and individual results.
- A breakdown of successful attacks by intent category (e.g., illegal acts, toxicity, violence) for targeted remediation.
- Clear mitigation actions for next steps.
Agentic Warfare™ Campaigns
Customers can leverage Agentic Warfare™ to execute advanced attacks based on custom intents, a dynamic approach to security testing that goes beyond signature-based attacks. Unlike pre-defined attack sets, Agentic Warfare attacks adapt to user-defined objectives, generating real-time adversarial prompts to uncover vulnerabilities in AI models. This method provides a more flexible and comprehensive way to assess AI security risks.
- Standard attacks – Signature-based, predefined attack sets.
- Agentic Warfare attacks – Dynamic, intent-based attacks generated from user-defined objectives.
Expanded Attack Arsenal – 22,000+ Signature Attacks
We’re released thousands of new signature attacks, including:
- 12,000+ new malicious prompts – Expanding our total attack set to over 22,000. We release new signature packs monthly.
- Persuasive Adversarial Prompts – Uses human-like persuasion techniques to subtly rephrase malicious intent.
Single Character Converter – A novel jailbreaking method that exploits vulnerabilities in short-length tokens.
CalypsoAI Security Index (CASI)
CASI is the industry’s first AI security scoring metric, helping organizations compare models on security—not just performance and cost. The CASI score (0-100) evaluates:
- Severity – The impact of a successful attack.
- Complexity – The sophistication of the attack.
- Defensive Breaking Point (DPB) – The weakest link in the model’s security defenses.
CASI is now embedded in Inference Red-team enabling customers to see the security of their AI systems in real-time.
Scoring Tiers:
- 0-69: Critical – Highly vulnerable; not production-ready.
- 70-85: Warning – Needs more safeguards before deployment.
- 85-99: Good – Secure against most attacks, but should be tested further.
- 100: Ideal – No vulnerabilities detected (verify with the latest attack signatures).
New Inference Defend Scanner Capabilities
Custom Regex & Keyword Scanners
We’ve added two powerful pattern-matching tools:
- Keyword Scanner – Block or audit specific terms (e.g., proprietary data, sensitive names).
- Regex Scanner – Define custom patterns for detection (e.g., email addresses, ID numbers).
Users can create and enable unlimited keyword and regex scanners.
New Financial & Medical Advice Scanners
To prevent AI systems from generating unauthorized advice, we’ve added two new scanners:
- Medical Advice Scanner – Blocks user prompts that seek direct medical guidance, prescriptions, or diagnoses without input from a licensed healthcare professional.
- Financial Advice Scanner – Blocks user prompts requesting personal financial recommendations, investment strategies, or money management advice.
These scanners help organizations comply with regulations, mitigate liability risks, and uphold user safety.
System Prompt & Obfuscation Scanners
As prompt injection attacks evolve, we’ve added dedicated scanners to our Prompt Injection Package:
- System Prompt Scanner – Detects attempts to expose system instructions.
- Obfuscation Scanner – Identifies text encoding tricks (e.g., base64, hex, leet speak, homoglyphs, intentional misspellings) used for adversarial attacks.
New Branding: Fresh Logo & UI Updates
We’re rebranding! You’ll notice a new CalypsoAI logo in our product UI, to go with the full website redesign.
Hugging Face Integration
With 1.4M+ models on Hugging Face, we’ve made it easier than ever to connect. Simply add a Hugging Face model using its name and API key.
CalypsoAI Platform Updates
Navigation Improvements
Our navigation bar was getting a little crowded, especially for laptop users. To optimize space, we’ve moved some elements to the user avatar menu in the top right, including:
- API Documentation
- Support
- Log Out
This update makes room in the left navigation for new features rolling out later this year.
