Security Report · raw.githubusercontent.com

HiddenLayer’s 2026 AI Threat Report Highlights Agentic Security and Supply Chain Gaps

Oct 21, 2026 · Original Report (PDF) ↗

Report discovered via Awesome Annual Security Reports.

Summary

HiddenLayer’s 2026 report examines AI supply chain threats, prompt injection and agent abuse. Its survey reports widespread shadow AI, inconsistent model scanning and limited AI-specific incident response and red teaming.

Key points

HiddenLayer reports that 76% of surveyed organizations consider unapproved or untracked AI deployments a definite or probable problem.
The survey says 93% use open-weight models, while fewer than half consistently scan incoming models for malicious content or integrity issues.
Only 29% report having a dedicated AI incident response plan, and 19% perform manual or automated AI red teaming.
The report describes indirect prompt injection and MCP tool poisoning that can redirect agent actions, expose credentials and abuse previously granted permissions.
It summarizes research showing that 250 malicious training documents could implant a trigger-based backdoor across tested model sizes; separate studies found that unlearning did not reliably remove poisoning.
The report covers AI-enabled extortion, LLM API credential theft and malware using language models, including PromptLock and LAMEHUG.
Recommended priorities include runtime monitoring, model and data provenance checks, enforceable tool-access controls and clearer accountability for AI security.

Article Details

Publisher: HiddenLayer
Scope: AI security survey findings, AI-enabled cybercrime, attacks against models and agentic systems, supply chain risks, and defensive practices, primarily covering 2025 and early 2026.
Sample Size: Survey respondent count not disclosed. Model-format analysis covers 15,325,540 files across 22 extension types, totaling 28.18 PB.
Key Statistics: 76% of surveyed organizations identified unapproved or untracked AI deployments as a definite or probable problem, compared with 61% in 2025.
93% of surveyed organizations use open-weight models; fewer than half consistently scan inbound models for malicious content or integrity issues.
Only 29% of surveyed organizations have a dedicated AI incident response plan.
Only 19% of surveyed organizations report performing manual or automated AI red teaming.
Unsafe pickle-based formats account for approximately 40% of model files in the report's model-format analysis.
Recommendations: Implement continuous AI monitoring, adversarial testing, runtime validation, and AI-specific incident response rather than relying on guardrails alone.
Monitor AI inputs, accessed resources, outputs, and behavioral changes, with enforceable runtime policies, anomaly detection, and kill-switch capabilities.
Maintain AI system inventories, risk classifications, traceability, documented controls, and evidence of control effectiveness.
Require AI vendors to explain how their production runtime controls monitor, detect, and respond to AI-specific threats.
Pin model dependencies to specific commits, verify models before cloning them into trusted internal storage, and scan codebases for hijackable model references and AI-specific configuration files.