Can Independent Testing Make AI Safer?

Vals AI CEO Rayan Krishnan said investment in AI testing and evaluation has not kept pace with rapidly improving model capabilities. The firm, which independently assesses models from OpenAI, Anthropic and others, sees early indications of recursive self-improvement, although current systems remain materially behind leading human researchers. The comments underscore both AI’s advancing capabilities and the growing need for robust model-evaluation infrastructure.

Analysis

The investable implication is not a near-term productivity inflection but a widening verification bottleneck. As model deployment moves into regulated and high-cost-error workflows, enterprise buyers will shift spend from experimentation toward testing, monitoring, provenance, and auditability. This favors application vendors with embedded workflow data and distribution—MSFT, NOW, CRM, ORCL, PLTR—over standalone model providers whose differentiation can erode as frontier capabilities converge.

The second-order beneficiary is cybersecurity. More autonomous agents expand identity, permissions, data-exfiltration, and software-supply-chain attack surfaces; PANW, CRWD, ZS, OKTA, and S are positioned to monetize incremental governance spend. The timing is likely 6-18 months, because meaningful budgets follow production deployments and procurement standards rather than model-demo cycles; near-term share-price impact is limited absent company-specific AI security bookings disclosures.

Consensus may be overpricing a direct leap from better models to lower software labor costs. Reliability improvements can initially raise enterprise IT spend through duplicated human review, red-teaming, insurance, and compliance processes, preserving demand for services and systems integration. ACN, IBM, and EPAM could therefore see a more durable transition revenue pool than the market expects, although their longer-run labor-arbitrage exposure remains negative if agent reliability reaches production-grade levels.

The thesis is falsified if enterprise AI pilots fail to convert into recurring production workloads, measured by slowing hyperscaler AI consumption growth, weak RPO commentary at NOW/CRM/ORCL, or cybersecurity vendors reporting no AI-related uplift in platform adoption. Conversely, a material model safety failure or regulatory enforcement event would accelerate governance spend while delaying broad application monetization.

Key Decisions for Investors

• No directional trade solely on this signal; establish a 1-3 month watchlist around PANW, CRWD, and OKTA for AI-governance bookings, net-retention, and platform-consolidation commentary. Upgrade only if management quantifies incremental AI-security demand rather than citing pilot activity.
• For a 6-18 month thematic position, prefer long PANW versus short a broad software proxy such as IGV: PANW has direct exposure to securing agentic workflows, while the software basket carries greater valuation risk if AI deployment raises implementation costs before it reduces headcount. Reassess if PANW billings growth decelerates materially relative to IGV or if AI-related demand remains unquantified through two earnings cycles.
• Maintain selective exposure to enterprise workflow owners NOW and ORCL rather than frontier-model proxies. Their installed bases can monetize evaluation and control layers through existing procurement channels; entry should follow evidence of AI attach-rate expansion or RPO acceleration, not model-capability headlines.
• Avoid adding to IT-services shorts on AI displacement narratives over the next 12 months. ACN and IBM may benefit from implementation, governance, and remediation work first; the bearish structural thesis requires demonstrated production autonomy and declining services headcount demand, neither of which is established by current evaluation commentary.