Anton Sidorov

I’m an AI security researcher and adversarial testing specialist focused on MCP (Model Context Protocol) and AI agent safety. My work is hands-on: I discover vulnerabilities in production MCP servers and run penetration testing against AI systems.

I’ve conducted security auditing of public MCP server infrastructure and uncovered 3 critical vulnerabilities. These include an authentication bypass enabling full tool exposure without credentials, prompt injection that manipulates agent context via tool outputs to execute unintended commands, and missing audit logging with no traceability for tool invocations, access patterns, or security events.

I actively research AI agent attack surfaces and mitigation strategies. I focus on prompt injection techniques targeting agent tool interfaces, and I’m building frameworks for systematic security assessment of AI agent workflows across multiple LLM backends.

I also contribute to AI safety research through analysis of the OpenAI Safety Bounty program. I map agent security concerns to bounty scope, develop proof-of-concept exploits for common vulnerability patterns, and share research insights to strengthen defenses for the broader AI security community.