Salt Security survey: 49.8% of organisations report unintended or unauthorised AI agent actions; shadow MCP servers rising
A Salt Security report, State of Agentic AI and API Security 2H 2026, finds that 49.8% of surveyed organisations confirmed or suspected an AI agent took an unintended, unexpected or unauthorised action in the past year. The research combines a survey of 380+ security and technology professionals with Salt Labs production telemetry. Among affected respondents, more than half saw an unauthorised external API call or transaction, and over 40% saw an agent modify or delete data. Visibility and containment are weak: only 12.5% can consistently trace an agent's full journey, and only 17.5% can automatically stop a misbehaving agent in production. The report also cites the recently remediated Manus agentic AI platform flaw, where a single malicious email could hijack the agent.
Detalhes técnicos
- 86.4% of respondents run at least one AI agent and 59.5% run more than ten. 76.8% have agents that chain actions across other agents, tools and APIs.
- 32.9% discovered an unauthorised, shadow or misconfigured Model Context Protocol (MCP) server in the past year, and 54.3% confirmed or suspected one. Salt Labs telemetry detected MCP infrastructure in 21.6% of customer environments in August 2026, up from 16.7% in January (roughly a 30% relative increase).
- 83.5% said AI-generated or AI-assisted code contributed to some application security findings. Only 6.8% resolve these issues within the coding assistant at the point of generation.
- In the Manus case, a single malicious email could hijack the platform and potentially expose the user's connected email, cloud storage and code repository accounts. It needed no stolen password and no link click. The guardrails detected the activity but could not stop the code from executing. The flaw was responsibly disclosed and fixed.
Remediação
The report recommends stronger visibility into agent activity, appropriate governance controls and automated containment capabilities for AI agents. Organisations should also inventory MCP infrastructure and catch AI-generated code issues earlier in development.
Linha do tempo
-
Salt Security survey: 49.8% of organisations report unintended or unauthorised AI agent actions; shadow MCP servers rising
ler a análise desta fonte ocultar
A Salt Security report, State of Agentic AI and API Security 2H 2026, finds that 49.8% of surveyed organisations confirmed or suspected an AI agent took an unintended, unexpected or unauthorised action in the past year. The research combines a survey of 380+ security and technology professionals with Salt Labs production telemetry. Among affected respondents, more than half saw an unauthorised external API call or transaction, and over 40% saw an agent modify or delete data. Visibility and containment are weak: only 12.5% can consistently trace an agent's full journey, and only 17.5% can automatically stop a misbehaving agent in production. The report also cites the recently remediated Manus agentic AI platform flaw, where a single malicious email could hijack the agent.
Technical Details
- 86.4% of respondents run at least one AI agent and 59.5% run more than ten. 76.8% have agents that chain actions across other agents, tools and APIs.
- 32.9% discovered an unauthorised, shadow or misconfigured Model Context Protocol (MCP) server in the past year, and 54.3% confirmed or suspected one. Salt Labs telemetry detected MCP infrastructure in 21.6% of customer environments in August 2026, up from 16.7% in January (roughly a 30% relative increase).
- 83.5% said AI-generated or AI-assisted code contributed to some application security findings. Only 6.8% resolve these issues within the coding assistant at the point of generation.
- In the Manus case, a single malicious email could hijack the platform and potentially expose the user's connected email, cloud storage and code repository accounts. It needed no stolen password and no link click. The guardrails detected the activity but could not stop the code from executing. The flaw was responsibly disclosed and fixed.
Remediation
The report recommends stronger visibility into agent activity, appropriate governance controls and automated containment capabilities for AI agents. Organisations should also inventory MCP infrastructure and catch AI-generated code issues earlier in development.