Securing Agentic Workflows: 1Password's Zero-Trust Strategy
1Password CTO Nancy Wang outlines a strategy for integrating AI coding agents into secure engineering pipelines. The discussion covers shifting security from checkpoints to runtime injection, measuring productivity via feature delivery rather than PR volume, and empowering non-engineers to build code. This approach reduces risk while accelerating development velocity.
The Shift to Runtime Security
The integration of AI coding agents has fundamentally altered the security landscape, moving the cost of development from code writing to code review and security auditing. 1Password CTO Nancy Wang argues that traditional security checkpoints create friction that slows down engineering. Instead, security must become the "paved path," embedded directly into the architecture. By injecting secrets at runtime rather than storing them on disk or in prompt contexts, organizations can enable agents to operate securely without exposing sensitive data. This "access without custody" model is critical for mitigating risks like AI-enabled ransomware, where stolen credentials can grant agents unauthorized system access.
Redefining Productivity Metrics
As AI makes code generation cheap, measuring productivity by the number of Pull Requests (PRs) or lines of code is no longer viable. Wang emphasizes that CTOs must shift focus to business outcomes, such as the number of features shipped or backlog items cleared. This approach ensures that increased velocity translates to tangible product value rather than just code volume. Additionally, organizations must maintain strict quality bars, including uptime and security standards, to ensure that AI-generated code does not introduce long-term technical debt or security vulnerabilities.
Expanding the Builder Role
A significant strategic shift involves empowering non-engineering roles, such as customer support representatives, to write and deploy code. By providing these teams with safe, sandboxed environments and coding agents, companies can resolve minor issues directly, bypassing the engineering backlog. This democratization of development requires a new approach to hiring and training, focusing on systems thinking and evaluation literacy rather than traditional coding tests. Ultimately, the goal is to create a culture where security and speed are not trade-offs but complementary drivers of innovation.
Key insights
-
Security should be embedded in the development workflow as a facilitator of speed rather than a checkpoint that causes delays. This involves injecting secrets at runtime and integrating security checks into CI/CD pipelines.
Impact: Reduces engineering friction and accelerates deployment cycles while maintaining high security standards.
-
AI agents should never hold raw secrets; instead, credentials should be injected only into authorized runtime processes. This 'access without custody' model significantly reduces the attack surface for AI-enabled threats.
Impact: Mitigates risks associated with prompt injection and data exfiltration by AI agents.
-
Productivity metrics must shift from code generation volume to feature delivery and business impact. Counting PRs is insufficient because AI can generate large volumes of low-value code.
Impact: Aligns engineering efforts with business goals and provides clearer ROI for AI tool investments.
-
Non-engineering staff, such as support reps, can be empowered to build and deploy code using safe, sandboxed environments. This reduces the engineering backlog and improves customer response times.
Impact: Unlocks hidden productivity and allows faster resolution of minor issues without engineering intervention.
-
Work should be classified based on risk and reversibility to determine the level of human oversight required. High-risk or irreversible changes must retain human-in-the-loop review.
Impact: Balances the speed of AI automation with the safety of human judgment for critical systems.
Action items
-
Implement runtime secret injection for AI agents to ensure they never have custody of raw credentials. Use tools that inject secrets directly into authorized processes rather than prompt contexts.
Impact: Enhances security posture against AI-driven attacks and complies with zero-trust principles.
-
Update engineering KPIs to focus on shipped features and backlog reduction rather than PR count or code volume. Establish clear definitions of 'value' for AI-generated work.
Impact: Provides a more accurate measure of AI tool effectiveness and aligns engineering with business outcomes.
-
Create safe, sandboxed development environments for non-engineering teams to allow them to fix minor bugs and deploy small features. Provide training on using coding agents within these constraints.
Impact: Reduces engineering backlog and empowers customer-facing teams to resolve issues faster.
-
Develop a classification system for engineering tasks that tags work based on risk, reversibility, and security impact. Define clear protocols for when human review is mandatory.
Impact: Ensures that AI automation is applied appropriately, maintaining safety for critical production systems.
-
Revise hiring and evaluation processes to assess systems thinking and evaluation literacy rather than traditional coding skills. Focus on candidates' ability to design robust agent workflows.
Impact: Builds a workforce capable of effectively managing and securing AI-driven development processes.
Quotes
“The best security tool, it's not a checkpoint. It actually removes things like weird workarounds. And it lets your engineers keep on working.”
“So we make sure is that the model itself never actually gets custody of the secret. And that eliminates your risk surface area.”
“So from a CTO perspective, right? how I measure whether these tools are effective isn't so much, hey, everybody went from generating, you know, 2.5 PRs a week to now 3.5.”