# Caleb Sima: AI Attackers Move at Machine Speed

## Executive summary

AI attackers operate at 'machine speed,' exhibiting extraordinary speed, thoroughness, and complexity by attacking all surfaces simultaneously. The discussion argues that traditional security controls, such as sandboxes, are insufficient against these advanced agents. To keep pace, defense must shift from human-led detection to fully automated, non-human-in-the-loop systems. Key defensive strategies include implementing forensic-grade detection, utilizing deception (digital twins/canaries), and automating remediation loops (e.g., autonomous patching) to manage the rapidly changing risk landscape.

## Key takeaways

- AI Attack Speed and Scope: AI agents can find 0-days by reversing open-source code and attacking multiple, disparate surfaces at once. For example, one report cited finding over 400 Linux 0-days from local user to root in an average of two and a half hours per exploit (00:05:52).
- Sandboxes are Insufficient Containment: Containing AI agents is extremely difficult. The speaker argues that external monitoring and external controls are critical because the ability for an agent to figure out how to break out of a sandbox is beyond current human reasoning capabilities (00:12:50).
- The Need for Automated Defense: Defenses must be automated to a high degree, eliminating the 'human in the loop' bottleneck. This includes implementing non-human-in-the-loop processes for tasks like instant, autonomous patching and remediation (00:44:10).
- Accountability Framework: In the event of an AI agent breaching a system and attacking a third party, the accountability rests with the user or company that invoked the agent, not the AI model itself (00:22:18).

## Technical details

- AI Agent Capabilities: Agents can perform complex, non-connected actions, such as taking credentials from an S3 bucket listing and brute-forcing them against a login page. They are capable of using stolen tokens and keys from other breached sites to maintain persistence (00:05:52).
- Detection Fidelity: Detection must evolve from low-fidelity capability (e.g., simple port scans) to extremely high-fidelity, forensic-grade monitoring. This requires tracking every file read/written and every socket opened to understand the attacker's pattern (00:36:47).
- Deception Techniques: Deception is crucial because traditional protection mechanisms are inadequate. Techniques include creating digital twins or 'mirror worlds' to funnel attackers into controlled environments (SE sheds, canaries, tripwires) for surveillance (00:36:47).
- Automated Remediation: The concept of 'one-way doors' (where a human might be needed) versus 'two-way doors' (like submitting a PR or making config changes) helps determine if a process can be fully automated by AI agents (00:48:53).

## Practical implications

- Implement automated remediation loops (e.g., autonomous patching) to eliminate the human bottleneck in the defense process.
- Shift security focus from prevention (which is failing) to detection and deception (luring attackers into monitored digital twins).
- Establish clear organizational agreements and frameworks that accept the risk of moving fast and mandate non-human-in-the-loop controls.
- Modern security teams must transition from merely reporting vulnerabilities to actively submitting and managing the fixes (PRs) themselves.

## Topics

AI Security, Cybersecurity, Automation, Agentic Systems, Software Factory, Tessl, Anthropic, OpenAI, Google Gemini

Source: https://www.youtube.com/watch?v=7NXdKekm95g