Hacking away at AI, ethically
When Betty Parman, Senior Manager for AI Security and Governance at a global manufacturer of high tech electronic devices, wanted to improve the security posture of its AI systems, she partnered with Infosys Responsible AI Office and Infosys BPM’s specialist expertise. This case details how Infosys BPM’s automated and manual red teaming exercises successfully identified and mitigated over 4,200 vulnerabilities, thus strengthening the systems’ overall resilience.
Approach summary:
- Assessment of AI architecture
- Prompt dataset curation
- Automated attack evaluation
- Manual attack evaluation
- Image guardrail testing
- Remediation of identified vulnerabilities
Key Benefits:
Executed 60,000+ prompt attacks across 9 risk areas using 40+ techniques 
Mitigated 4,200+ AI vulnerabilities (~70 per 1K prompts) 
Uncovered 210 high risk issues via 16+ manual attack methods complementing automation (~66% discovery rate) 
Delivered 10+ stakeholder reports with risks and remediation insights 
Built a repeatable AI risk framework combining automation, red teaming, and image guardrails 
