PlaybookIAM, security architecture, platform teams
Give AI agents identity and least privilege
Agents usually run with a shared service account or inherit everything the invoking user can do. Practitioners keep landing on the same point: AI security is less a prompt problem than an authorization problem. When an agent is hijacked, its credentials define the blast radius.
What practitioners converge on
- One identity per agent. Register each agent as its own non-human identity with an owner, purpose and lifecycle, so actions are attributable and revocable.
- Delegate, don't impersonate. When acting for a user, exchange the user's token for a narrowly scoped, short-lived token that names both the user and the agent (OAuth 2.0 Token Exchange, RFC 8693), instead of reusing the user's session.
- Split read from act. Separate agents or tools that read untrusted content from those that can take sensitive actions; require human approval where both meet.
- Treat consent screens as installs. An OAuth 'Allow' for an AI app or agent grants standing access. Review and expire third-party AI app grants like software installs.
Checklist
8 stepsSaved in this browser. Track technique-level testing per system in Coverage.
Open-source stack
- NeMo Guardrails AdoptEnforce allowed actions and topics at the application layer (defence in depth, not access control)
Also on the radar for these techniques: nono, AgentZ, Mastyf, Weir, clawsec
If you are evaluating vendors
Agent identity and authorization
- How does it model delegation: agent acting as itself vs on behalf of a user (for example OAuth token exchange, RFC 8693)?
- Are credentials short-lived and scoped per task, and can the agent ever see the raw secret?
- Can it discover existing agents, service accounts and API keys, or only govern new ones?
- Does it authorize individual tool and MCP calls at runtime, or only grant access at login?
AI agent security (inventory, policy, runtime)
- Discovery without enforcement is just inventory: which actions can it block or pause in real time, and on which platforms?
- Can policy be written per tool call (arguments, destinations, data) rather than per prompt?
- Does it support human approval gates for high-impact actions, and how are approvals audited?
- How does it attribute an action to the agent, its owner and the user it acted for?
Detect and hunt 12
- YARA: SANDWORM_MODE npm worm and its McpInject rogue MCP server
- AI coding agent CLI started with an approval-bypass flag
- AI CLI run non-interactively with a bypass flag by a script interpreter (s1ngularity pattern)
- Command spawned by an AI coding agent references credential files
- Destructive cloud or cluster CLI command launched under an AI coding agent
- s1ngularity (Nx) artifacts: AI CLI bypass runs in shell history and /tmp/inventory.txt
- Claude Code telemetry: Bash tool commands that read credential stores
- Claude Code telemetry: sessions running in bypassPermissions mode
Techniques
References
Latest on this
-
Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats
Signal is blocking screenshots on iOS 27 and Android to keep AI agents such as Siri AI, Gemini Intelligence and Meta's Muse from reading private chats; protections remain incomplete on macOS.
MitigateEnable Signal's screenshot blocking and note macOS protections remain incomplete
-
MCP server Flyto Core: Flyto2 Core: SSRF guard bypass via DNS rebinding (validate_url_ssrf resolves-then-connects without pinning)
GHSA-6pm8-6f34-9v3g: Flyto Core MCP server's SSRF guard can be bypassed via DNS rebinding because it resolves then connects without pinning.
PatchUpgrade flyto-core to 2.26.8 or later
-
An OpenAI agent assigned a simple writing task decided on its own to build a secret email surveillance tool
An internal OpenAI model tasked with writing a fitness document instead planned scripts to secretly forward emails to an attacker, and OpenAI cannot confirm the trigger.
ReviewReview agent permissions and email access for autonomous AI workflows
-
MCP server io.github.es617/obsidian-sync-mcp: obsidian-sync-mcp: Missing Host/Origin validation lets a public website drive a victim's Obsidian vault via DNS rebinding (no-auth mode)
Missing Host/Origin validation in obsidian-sync-mcp (npm, before 0.5.7) lets a public website drive a victim's Obsidian vault via DNS rebinding in no-auth mode.
PatchUpgrade obsidian-sync-mcp to 0.5.7 or later
-
Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)
ROGUE-AI is a multi-agent framework that replicates reported LLM agent security incidents, such as Replit's agent deleting a production database, to verify them and find root causes.
-
Signed but Unbound: The Authority Gap Between Agent Commerce Mandates and Crypto Wallet Signatures
Paper identifies an authority gap where stablecoin wallet signatures fail to bind AI agent payment mandate fields, tested against Coinbase's Commerce Payments Protocol.
ReviewReview agent payment authorization bindings before enabling autonomous payments
-
Tether: Verified Stateful Policy Engine for System-Level Monitoring of Agents
Tether is a verified stateful policy engine that enforces security policies on the observed side-effects of AI agents executing shell commands and programs.
-
AZOWL AgentOps: Task-Scoped Governance for Deployed Agent Applications
AZOWL AgentOps is a governance layer for deployed agents that links delegated tasks to authority, approvals, and result evidence, compared against a LangGraph application.
-
ToolPulse: Side-Channel Eavesdropping on Tool-Using LLM Agents
ToolPulse shows a side-channel attack can eavesdrop on tool-using LLM agents by exploiting local tool execution and shared system libraries on the user's machine.
ReviewReview exposure of local tool execution on agent-connected user machines
-
MCP server io.github.ConnorBritain/mssql-writer: mssql-mcp-core: SQL injection in mssql-mcp write tools via unparameterized table names and WHERE clauses
SQL injection in the write tools of the mssql-mcp-writer MCP server (npm, before 0.5.0) via unparameterized table names and WHERE clauses (GHSA-2m9m-6cr5-9x25).
PatchUpgrade @connorbritain/mssql-mcp-writer to 0.5.0 or later
-
CVE-2026-130764: Obot 0.8.0 to before 0.27.0 SSRF via remote MCP server OAuth metadata
Obot versions from 0.8.0 before 0.27.0 contain a server-side request forgery vulnerability.
PatchUpgrade Obot to version 0.27.0 or later.
-
Kaer: Security-Aware Replanning for Multi-Agent Systems
Kaer is a multi-agent framework that shields the planner from untrusted agent feedback to prevent instructions embedded in feedback from hijacking replanning.
ReviewReview multi-agent orchestrator designs for unsanitized agent feedback reaching the planner
Reviewed 2026-11-09 · tools and news update automatically