EnkryptAI Rating
EnkryptAI Rating for Model Safety that takes into account model vulnerabilities as well as the severity of those vulnerabilities.
1.7
NIST Score
Denotes the percentage of number of attacks successful against total number of atacks performed.
41%
OWASP Score
Weighted average of all the successful attack percentages according to risk ranking of OWASP.
49%
G
gemma-2-9b-it
Provider: Google
Rank : 177/234
Bias Bias Rating is calculated by inversely mapping the Bias risk score for all models (ranging from 0 to 100) onto a rating scale from 5 (safe) to 0 (risky). The bias risk score of each model is the percentage of successful bias attacks. | Harmful Tests Harmful Tests Rating is calculated by inversely mapping the Harmful Test risk score for all models (ranging from 0 to 100) onto a rating scale from 5 (safe) to 0 (risky). The Harmful Test risk score is the percentage of successful Harmful Tests. | Toxicity Toxicity Rating is calculated by inversely mapping the Toxicity risk score for all models (raning from 0 to 100) onto a rating scale from 5 (safe) to 0 (risky). Toxicity Risk score is the percentage of successful toxicity generation attacks. | CBRN CBRN Rating is calculated by inversely mapping the CBRN risk score for all models (raning from 0 to 100) onto a rating scale from 5 (safe) to 0 (risky). The CBRN Risk score for each model is the percentage of successful attacks for Chemical, Biological and Cybersecurity information generation. | Insecure Code Insecure Code Rating is calculated by inversely mapping the Insecure risk score for all models (raning from 0 to 100) onto a rating scale from 5 (safe) to 0 (risky). The Insecure Code risk for each model is the percentage of successful attacks. | Robustness Tests Robustness Rating is calculated by mapping the Robustness score for all models (raning from 0 to 1) onto a rating scale from 0 (risky) to 5 (safe). The Robustness Score indicates how a robust is the model to input variation. | |
|---|---|---|---|---|---|---|
2.6 | 3.7 | 0.2 | 4.6 | 0.9 | TBD | |
 | Harmful Bias & Homogenisation | Dangerous, Violent, or Hateful Content | Obscene, Degrading, and/or Abusive Content | CBRN Information or Capabilities | Information Security | Confabulation |
 | LLM01 Prompt Injection LLM04 Data and Model Poisoning | LLM01 Prompt Injection | LLM01 Prompt Injection | LLM01 Prompt Injection | LLM01 Prompt Injection LLM05 Improper Output Handling |
Compare Models