VentureBeat published two Pulse Research surveys on July 16 that together paint a grim picture: AI agents are running far faster than enterprises are prepared for.
Survey 1: The Agent Security Gap (107 enterprises)
| Key figure | Value | |---|---| | Agent security incidents or near-misses | 54% (18% confirmed + 36% near-miss) | | Agents with scoped, independent managed identities | 32% | | Agents sharing credentials (API keys / borrowed human credentials) | 69% (multi-select) | | High-risk agents sandboxed | 30% | | Monitoring/logging agent activity | 47% | | Enforced scoped runtime permissions | 49% | | Vendor-native security stack (OpenAI guardrails) | 51% | | Average satisfaction | 4.2 / 5 |
Notably, larger enterprises get hit more: incident rates are 49% for 100–1,000 employee companies versus 63% for 1,000+—yet sandbox adoption at 1,000+ employee companies drops from 35% to 20%.
Money is moving: Palo Alto Networks, CrowdStrike, and Cisco have collectively committed $22 billion toward AI agent security within 12 months (Palo Alto's $21.1B CyberArk acquisition, CrowdStrike's $740M SGNL deal, Cisco's $400M Astrix Security purchase).
Survey 2: The Agent Evaluation Gap (157 enterprises)
| Key figure | Value | |---|---| | Deployed agents in the past 12 months that passed internal evals but failed customers in production | 50% (24% multiple times) | | Fully trust automated evaluations | 5% | | Cite "poor alignment between evals and real outcomes" as the top gap | 29% | | Allow or engineering fully automated, zero-human deployment | 66% (34% already + 32% building) | | Perform real-time quality checks on actual agent output | 23% | | Primary tooling | Vendor-native evals and "no dedicated tool" tied at 17% | | Large enterprises (2,500+) moving fastest to zero-human deployment | 70% |
Among companies with actual incidents, large enterprises (2,500+) hit 54% versus 48% for smaller firms—the biggest companies move most aggressively yet still fail to catch failures.
Vendor positioning
- Palo Alto Networks closed its $21.1B CyberArk acquisition on Feb 11 (announced at $25B last July, the largest deal in company history).
- CrowdStrike acquired runtime authorization platform SGNL for $740M and launched Continuous Identity for AI Agents on June 15.
- Cisco announced its ~$400M acquisition of non-human identity specialist Astrix Security on May 4.
- Both surveys are single-point June snapshots with self-selected samples skewed toward mid-market—treat as the perspective of companies actively building agent security, not the largest operators.
- 5% fully trusting automated evals doesn't mean only 5% trust any evaluation—manual review may be serving as a backstop; VentureBeat doesn't break this out.
- 66% "automated deployment" may use loose definitions of low-risk agents; follow-up breakdowns are needed.
- VentureBeat — the agent security gap: https://venturebeat.com/ai/the-agent-security-gap-54-of-enterprises-have-already-had-an-ai-agent-incident-and-most-still-let-agents-share-credentials
- VentureBeat — the agent evaluation gap: https://venturebeat.com/ai/the-agent-evaluation-gap-enterprise-ai-organizations-have-a-reality-alignment-problem-not-a-coverage-problem-and-most-are-shipping-to-production-anyway
- Secondary analysis: https://valueaddvc.com/pulse/shared-api-keys-ai-agent-security-2026 and https://saassentinel.com/2026/07/12/ai-agents-are-outpacing-enterprise-ability-to-verify-them
Analysis
"Passing evals = ready for production" is a broken assumption
Traditional evals (unit, integration, end-to-end testing) work for systems with bounded inputs/outputs and reproducible errors. Agents are not that: behavior depends on prompts, user context, and environment state jointly. Automated evals can, at best, certify performance on a fixed test set. As VentureBeat puts it: "a passing eval is not the same as a working agent." With 50% of enterprises having shipped eval-passing agents that failed customers, evaluation-as-quality-gate is failing, and 29% cite eval-to-reality misalignment as the top problem.
Zero-human deployment: nobody trusts evals, yet 66% rely on them
The core contradiction: only 5% fully trust automated evaluations, but 66% use them as the automatic gate for production deployment—a guardrail that itself lacks certification. Warning sign: the fastest movers toward zero-human deployment are 2,500+ employee enterprises (70%), whose governance teams may be least familiar with novel agent failure modes.
Shared credentials amplify incidents
The cascade is clear: 54% incident rate + 69% shared credentials + only 30% sandboxing means a single compromised agent can spread damage across every workflow sharing its keys, with 70% of enterprises lacking a sandbox backstop. Machine identities already outnumber human ones at 82:1 per CyberArk research, and most IAM systems were never designed for that ratio.
Where the money points
$22B in M&A vastly exceeds current agent-security revenue—this is a bet that agent security becomes the next IAM battleground of the cloud era. For startups, the clear openings are agent identity, agent scoped credentials, and agent sandboxing, with a 12–24 month window before incumbents consolidate.
Why it matters
1. Passing evals is no longer a quality gate—pre-deployment evaluation plus real-time production verification must both be in place. 2. Shared credentials are the easiest hole to poke—if your agents use the company's shared API key, revisit your IAM policy. 3. Agent security is the next cloud security battlefield—the acquisition prices show big vendors have committed; the startup window is 12–24 months. 4. The evaluation gap is the key AI-governance theme for H2 2026—large-scale enterprise rollouts like ChatGPT Work and Claude Code will worsen it.
Caveats
Sources: