AI agents are reportedly making it easier for attackers to find and exploit security vulnerabilities in software code. This development has led to a focus on new methods for software engineering and application security teams to identify and resolve these weaknesses.
October 2026 saw incidents highlighting these capabilities. Pillar’s OpenCode disclosure revealed how a malicious Model Context Protocol (MCP) server could execute commands on a developer’s Windows or WSL machine during authentication. Additionally, the discovery of two PHP2Shell remote code execution bugs in the PHP runtime, aided by OpenAI models, underscored the urgency of securing existing code.
Microsoft also reported an instance where an AI agent progressed from searching GitHub for exposed credentials to stealing Azure Key Vault data within five minutes. The agent reportedly corrected API versions and retried requests until successful.
A report by Anthropic described individual operators using agent frameworks for multi-victim campaigns, which previously required specialist teams. These agents carried out reconnaissance, exploitation, and data theft across multiple targets in parallel, making sophisticated operations more accessible.
These new agentic attacks suggest that teams should re-evaluate their existing code. Traditional application security checks, such as rule-based static analysis, dependency scanning, and secret detection, may no longer be sufficient on their own. Developers, engineering leaders, and security teams are now encouraged to use defensive AI agents to trace attack paths, uncover missing safeguards and business-logic vulnerabilities, and assist with fixes.