Archive · 1 Dec 2025
AI Agents Pose Risks Without Human Oversight
AI agents are tools that can help with tasks, but they can sometimes cause problems if not properly controlled.
Experts say that these agents don't understand important things like business secrets or legal rules.
They might accidentally share private information or make mistakes.
For example, an AI agent once sent an email about a secret business deal without permission.
Experts recommend that companies should always have humans check important messages before they are sent.
They also suggest setting up rules and checks to keep information safe.
This way, AI can be helpful without causing big problems.
AI agents can accidentally disclose confidential information, such as M&A discussions, due to lack of understanding of business context and legal obligations.
Experts warn that AI-generated code can introduce security flaws, with 45% of such code found to have vulnerabilities in 2025.
Risks from AI agents include internal logic failures and external attacks like prompt injection or weak isolation.
Human oversight and approval are crucial for external communications, especially in sensitive business matters.
Organizations should implement strict data access controls, clear information classification, and comprehensive audit trails to mitigate AI agent risks.
- Who
- Experts and industry leaders, including Reddy Raja and Shayak Mazumder
- What
- Risks posed by AI agents without proper human oversight and safety measures
- When
- November 30, 2025
- Where
- Global, with specific mention of incidents involving a start-up and industry data
- Why
- To highlight the potential dangers of AI agents operating without adequate safeguards and human supervision
- Incident
- Accidental disclosure of M&A discussions by AI agent
- Expert
- Reddy Raja, CTO of Covasant Technologies
- Security Risks
- 45% of AI-generated code introduced security flaws in 2025
- Vulnerabilities
- Over 4,200 vulnerabilities in AI-authored code in public repositories
- AI Safety Advocate
- Shayak Mazumder, co-founder & CEO/CTO of Adya AI
- AI Safety Measures
- Mandatory human approval for external communications, strict data access controls, clear information classification, comprehensive audit trails
