Report: 54% of Enterprises Hit by AI Agent Security Incidents
VentureBeat AI · rss · 2026-07-17
A VentureBeat survey of over a hundred mid-to-large enterprises reveals a stark "agent security gap": **54% of surveyed companies have experienced AI agent security incidents or near-misses** (18% confirmed leaks, 36% proactively intercepted). **The core vulnerabilities lie in identity and access management:** - **Rampant credential sharing:** Only 32% of enterprises provision unique, restricted identities for each agent. A massive 69% share API keys or human account credentials among agents. Data shows that companies with credential sharing have a significantly higher incident rate (63.5%) compared to those enforcing strict unique identities (40.9%). - **Lack of isolation:** Only 30% of enterprises run high-risk agents in sandboxes. Larger enterprises see higher incident rates, yet their adoption of sandbox isolation is paradoxically lower. **Defense tools and awareness:** - Corporate defenses rely heavily on native tools from major providers (like OpenAI guardrails or Microsoft cloud controls), with dedicated agent security tools almost entirely absent. - Although average satisfaction with current security stacks sits high at 4.2/5, only a third believe their defenses outpace AI attackers, and most plan to replace their existing security tools within the year.
More from coding & agent
- Three retrieval models fine-tuned on synthetic agent profiles generalize to unseen marketplaces — _reachsumit · 2026-07-21
- Deep research agents lose 66–88 points in misleading-evidence benchmarks — _reachsumit · 2026-07-21
- RIMS uses LogSumExp aggregation to make small RAG models more robust to noisy evidence — _reachsumit · 2026-07-21
- Karpathy Challenges AI Assumption: Better Agents Don't Need Bigger Models — Div_pradeep · 2026-07-21
- An article breaks down how context compaction works in AI systems — blaizedsouza · 2026-07-21
- Chaos engineering surfaces failure modes in production agent systems — blaizedsouza · 2026-07-21