- Enterprises give AI agents the run of their code and production systems, with little ability to see or stop them
- Fixing a code flaw takes 18.4 days on average, but frontier AI can find and exploit flaws in minutes
- Telcos run the tightest pipelines of any sector, give agents the most latitude and worry the least
Organizations are YOLO-ing agentic AI, giving agents broad powers to modify enterprise code and production systems while failing to provide adequate supervision, according to a study by Palo Alto Networks released Monday.
Nearly three-quarters (73%) of organizations let AI agents modify code, pipeline configurations and production environments. And yet many organizations don't know what agents are doing, lack the ability to tell agents' work apart from humans' and can't cut off agentic access when necessary, according to Palo Alto Networks' State of Cloud Security Report 2026.
"Traditional security controls, fragmented visibility and manual response processes were not designed for environments where autonomous agents can make changes at machine speed," wrote Sarit Tager, VP of product management at Palo Alto Networks, in a blog post about the report.
While enterprises dawdle, attackers are speeding up. Frontier AI is giving adversaries the ability to find vulnerabilities, build exploits and launch sophisticated attacks faster than ever, Tager wrote. Indeed, Palo Alto's own agentic research system, NOVA, found 14,090 unreported vulnerabilities across 3,915 open-source projects in about two months.
Whose name is on the work?
More than half (53%) of organizations log agent activity under the human developer's identity or personal access token. In the logs, the agent's work looks like the engineer's. And 40% of organizations can't fully reconstruct what an agent did after the fact.
These limitations make it difficult for enterprises to identify and mitigate problems. Investigators need to know which systems an agent touched and whether it strayed beyond its assigned task, so they can decide what to contain and what to roll back. Without a distinct identity and an audit trail for each agent, investigation and containment slow down, according to the report.
More problems
Revoking or rotating compromised credentials takes 2.5 days on average, and 60% of organizations still rely on manual intervention to revoke nonhuman access. Only 9% can automatically discover and classify ungoverned AI running in their cloud environments.
Fixing a security flaw in code takes 18.4 days on average, from detection to deployment in production. That was bad enough when bugs were found at human speed, but frontier models can find and weaponize flaws in minutes, leaving an organization exposed long after a vulnerability becomes discoverable, according to the report.
Catching up is expensive. Bank of America brought in about 1,000 contractors to patch flaws exposed by AI models, including Anthropic's Mythos.
Good news for telcos — for a change
The 205 telecom respondents ran the tightest pipelines of any sector. They led in enforcing infrastructure-as-code scanning before production (41% vs. 34% globally) and in revoking compromised credentials within one business day (49% vs. 38%).
Should telcos be more worried? The data suggests so. Some 35% let AI agents modify CI/CD or infrastructure-as-code configurations, tied with technology companies for the highest share (global: 27%). Only 66% can trace AI-generated code from commit to cloud runtime, compared with 75% globally. Yet just 48% rated frontier AI threats a significant or severe danger to critical infrastructure, the lowest of any sector, compared with 65% overall.
That's a lot of calm for operators of critical infrastructure who give agents more authority than nearly anyone else.
Palo Alto's advice
The report draws on a survey of 2,800 security leaders and practitioners across 10 countries and 13 industries, fielded between June 16 and July 8. The data is self-reported.
Palo Alto's recommendations include distinct identities for agents, short-lived credentials, automated revocation and unified security context from code to cloud to SOC — all of which line up with products it sells. The company has been pursuing the role of enterprise security control plane through an acquisition spree worth more than $35 billion.
Learn more about Palo Alto Networks and AI agent security on Fierce Network
Palo Alto: AI broke the patch window — We can fix it
Opinion: Security vies to become the new control plane of the enterprise
Identity token management is critical for CSPs in the AI era
Agentic AI intros 5 new risks for telcos — no one is talking about them