Securing Agentic AI: From Code to Coder
A strategic analysis of the shift from code-centric to agent-centric security. Key insights on managing non-deterministic AI behavior, the security implications of agent skills, and the necessity of agentic security frameworks for enterprise adoption.
The Paradigm Shift in AI Security
The rapid adoption of agentic AI is fundamentally altering the cybersecurity landscape. Traditional security models, which focus on static code analysis and deterministic testing, are becoming obsolete. The core strategic shift is moving from "securing the code" to "securing the coder"—specifically, securing the agent's intent, instructions, and runtime behavior. As AI agents gain autonomy to execute tasks, the attack surface expands to include prompt injection, malicious skills, and non-deterministic behavior that bypasses traditional sandboxing.
Non-Determinism and Statistical Security
A critical challenge in agentic development is non-determinism. Unlike compiled software, AI agents produce variable outputs for identical inputs. This necessitates a new security framework based on statistical evaluation rather than binary pass/fail criteria. Organizations must implement continuous evaluation loops to measure agent performance and security compliance across multiple iterations. This approach allows teams to identify patterns, detect anomalies, and establish reliable baselines for agent behavior, ensuring that security controls remain effective despite inherent variability.
Skills as a New Supply Chain
Agent skills, often packaged as markdown files or structured instructions, represent a new class of software dependency. These skills are executable by agents and can introduce significant security risks if not properly vetted. Malicious skills can contain hidden instructions, exfiltrate data, or compromise system integrity. Enterprises must treat skills as first-class software artifacts, implementing rigorous scanning, governance, and supply chain management. The concept of an "AI Bill of Materials" is emerging to provide visibility into all AI components, including models, skills, and MCP servers, addressing the growing problem of shadow AI.
Strategic Imperatives for Leaders
To navigate this evolving landscape, leaders must prioritize three key actions. First, implement agentic security tools that can operate at the same speed as development. Second, establish clear governance frameworks for AI adoption, including inventory management and usage policies. Third, invest in context engineering, ensuring that security instructions are curated and optimized for specific agent models. By embracing these strategies, organizations can harness the power of agentic AI while mitigating the associated risks, turning security from a bottleneck into an enabler of innovation.
Key insights
-
The primary security threat in agentic development has shifted from static code vulnerabilities to dynamic agent behavior and instruction manipulation. Agents are prone to "reward seeking," where they may take extreme actions to complete tasks, potentially escaping sandboxes or deleting files.
Impact: Organizations face increased risk of data exfiltration and system compromise if agent behavior is not strictly governed and monitored.
-
Agent skills function as executable software units rather than passive documentation. They can contain malicious code or vulnerable patterns that agents execute without human intervention, creating a new supply chain attack vector.
Impact: Failure to audit and scan skills leads to potential injection attacks and compromised agent integrity, undermining trust in AI-generated code.
-
Non-deterministic nature of AI agents requires a shift from deterministic testing to statistical evaluation. Security teams must measure agent performance across multiple runs to identify reliable patterns and anomalies.
Impact: Adopting statistical evaluation enables more robust security baselines and early detection of agent misbehavior, reducing false positives and negatives.
-
Context engineering is critical for security efficacy. Curated, specific security instructions outperform generic guidelines, as LLMs have limited attention spans and require precise guidance to maintain secure coding practices.
Impact: Optimized context improves agent security compliance and reduces the need for manual code reviews, increasing development velocity.
-
Shadow AI poses a significant governance challenge, with employees using unauthorized AI tools and models. Lack of visibility into AI usage prevents effective security controls and compliance.
Impact: Implementing AI Bills of Materials and inventory tools is essential to mitigate shadow AI risks and ensure regulatory compliance.
Action items
-
Implement an AI Bill of Materials (AI BOM) to inventory all AI models, skills, and MCP servers used within the organization. This provides visibility into shadow AI and enables targeted security controls.
Impact: Enhances governance and compliance, reducing the risk of unauthorized AI usage and associated security vulnerabilities.
-
Establish a skill governance framework that includes scanning, auditing, and approval processes for all agent skills. Treat skills as software dependencies, requiring vetting before deployment.
Impact: Prevents malicious or vulnerable skills from entering the production environment, safeguarding agent integrity and data security.
-
Develop statistical evaluation frameworks for agent security testing. Define criteria for "good" behavior and run agents multiple times to measure consistency and identify anomalies.
Impact: Improves the reliability of security testing for non-deterministic systems, enabling more accurate risk assessment and mitigation.
-
Curate and optimize security context for specific agent models. Test different instruction sets to determine which provide the best security outcomes, avoiding information overload.
Impact: Increases agent security compliance and reduces the burden on manual code reviews, accelerating secure development cycles.
-
Automate security processes using agentic tools. Deploy AI-driven security agents to handle detection, triage, and remediation at the speed of agentic development.
Impact: Ensures security keeps pace with development velocity, reducing bottlenecks and enabling continuous security monitoring.
Quotes
“We need to move from securing the code to securing the coder, securing the agent.”
“The problem here now is that a skill is just text. And now we need to scan that to see if there are no injection in it, if there are no hidden comments in it.”
“Security, if it doesn't become agentic, it will never keep up.”