Security News

AI Agents Make Attacking Your Code Cheaper and Faster, Forcing a Rethink of Application Security

The Register · 17 Nov 2026
Key Takeaway Review your older, already-deployed code for missing permission checks and business-logic gaps, not just new changes, because cheap AI agents can probe all of it.

A partner article on The Register argues that October 2026 may be remembered as a turning point for software developers and application security teams. Its central claim is that AI agents can now investigate unfamiliar systems, test ways in and change tactics when an attempt fails, and that anyone with a credit card can use this against an application.

The article points to several examples. Pillar's OpenCode disclosure showed how a malicious Model Context Protocol (MCP) server could run commands on a developer's Windows or WSL machine during authentication, with a server-supplied URL reaching a shell. Two PHP2Shell remote code execution bugs in the PHP runtime were found with help from OpenAI's models. Microsoft reported an agent that went from searching GitHub for exposed credentials to stealing Azure Key Vault data in five minutes, correcting its API versions after failed requests. Anthropic described individual operators running multi-victim campaigns that once needed specialist teams.

Cost is a major factor. Analyst Tomasz Tunguz estimated a 100-million-token attack workload on a low-cost model at $90.20, against the $5,000 to $10,000 frontier-model estimate cited by the Bank for International Settlements. That figure covers model usage only, not the full cost of an attack.

The article says rule-based static analysis, dependency scanning and secret detection are no longer enough alone. It argues for defensive AI agents that trace attack paths and find missing safeguards and business-logic flaws, including in code that has run for years. It promotes CodeRabbit Secure, which offers PR security reviews, repository-wide Deep Scans, and merge blocking on GitHub. As a sponsored piece, its product claims are the vendor's own.

Why it matters

Controls that look for known patterns, such as rule-based scanning, dependency checks and secret detection, will not catch a missing permission check or a flawed business rule, which is what an adaptive agent can hunt for in code that has passed every scan. The falling cost of attacks also weakens the assumption that a small codebase is not worth targeting. Teams must now decide how to review long-running production code, not only new pull requests.

What to do

  • Review older production code paths for missing permission checks and business-rule enforcement, not just the latest PR.
  • Check developer machines and tooling for untrusted MCP servers, given the OpenCode authentication command execution issue.
  • Search your repositories for exposed credentials before an attacker does, as in the GitHub to Azure Key Vault case.
  • Require security checks on pull requests through branch protection so unresolved high-severity findings block merges.
Running AI agents? Eight controls for securing them, mapped to the ISM ->

Summarised by CISO AI from The Register, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.