PlaybookSecurity engineering, IT, developer platform teams
Secure AI coding tools in your organisation
Developers run Claude Code, Cursor, Copilot, Codex and similar agents with access to source, terminals, cloud credentials and MCP servers. These tools keep local transcripts (including anything pasted, such as secrets), send context to the vendor, and can be steered by instructions hidden in repositories, issues or web pages.
What practitioners converge on
- Pin a minimum safe version. AI coding tools ship security fixes often. Track advisories per tool and enforce a minimum version through MDM or package policy.
- Managed settings over guidance. Use the vendor's admin or managed-settings mechanism to set permissions, disable risky auto-run modes, restrict MCP servers and control data retention.
- Assume transcripts contain secrets. Local history and logs can hold pasted credentials and file contents. Include their paths in secret scanning, endpoint DLP and offboarding.
- Untrusted repo = untrusted prompt. Project config files, READMEs, issues and dependencies can carry instructions. Open untrusted code without agent mode or auto-approve.
Checklist
8 stepsSaved in this browser. Track technique-level testing per system in Coverage.
Open-source stack
- nono AssessKernel-enforced sandbox to limit what a coding agent can read, write and reach
- mcp-scan TrialScan MCP configs on developer machines for poisoned or risky servers
- promptfoo AdoptTest internal coding agents and their tool permissions
Also on the radar for these techniques: Cisco MCP Scanner, AgentZ, Mastyf, Weir, clawsec
Tools covered
Current advisories, minimum safe version, local data and hardening settings for each tool.
- Claude Code ≥ 2.1.251
- Cursor ≥ 3.1.2
- GitHub Copilot
- OpenAI Codex CLI ≥ 0.39.0
- Gemini CLI ≥ 0.39.1
- Windsurf
If you are evaluating vendors
AI coding agent security
- Which coding agents and IDEs are supported today, and on which OSes?
- Can it block specific commands (rm -rf, force push, prod DB writes) before they run, or only log them?
- How does it stop secrets in files, env vars and transcripts from reaching the model?
- Does it vet MCP servers, skills and extensions before install, and detect changes later?
AI usage control (Shadow AI, GenAI DLP)
- Is this enforcement or only discovery? Show a prompt with a secret being blocked or redacted inline, not an alert afterwards.
- Which surfaces are covered today: web apps, desktop apps, IDE plugins, CLIs, agentic browsers, embedded AI inside SaaS?
- How do you handle personal accounts on sanctioned tools (personal ChatGPT vs the enterprise tenant)?
- What is the false-positive rate on our own prompts and files during the POC, and who tunes the classifiers?
MCP security (gateways, scanning, registries)
- How does it find MCP servers that developers configured locally (stdio) and never routed through the gateway?
- Does it detect tool poisoning and rug pulls, such as tool descriptions that change after approval or contain invisible Unicode?
- Can policy be set per tool and per argument, not just per server?
- How are downstream credentials handled: does the gateway hold tokens, or do agents still see raw secrets?
Detect and hunt 23
- YARA: s1ngularity (Nx) telemetry.js payload and shell-profile shutdown line
- Claude Code / VS Code project persistence files from the Mini Shai-Hulud worm
- 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
- AI coding agent spawns a shell that pipes a download into an interpreter
- Destructive cloud or cluster CLI command launched under an AI coding agent
Techniques
References
Latest on this
-
OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms
Dragos reported attackers stole data from three AI inference providers for months, including requests flagged Zero Data Retention; OpenRouter dropped the three providers.
ReviewReview data retention settings and provider routing used with AI inference gateways
-
A Single Malicious Webpage Let Researchers Hijack Meta Muse and Access Private Chats
Researchers showed a malicious webpage could extract a Meta Muse user's private chats and stored memories without an approval prompt, in a lab demonstration.
ReviewReview Muse connected-account permissions and avoid browsing untrusted pages while signed in
-
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
-
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
-
Ask in the Crowd: Differentially Private LLM Inference via Dummy-Augmented Shuffling
DP-SPD is a differentially private LLM inference framework that hides real user queries among client-generated dummy queries before sending them to remote LLM services.
-
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.
-
Why the Harness Matters: Early Lessons From the Evo Code Scanning Agent Preview
Snyk published early preview results for its Evo Code Scanning Agent, reporting it found more known vulnerabilities than Claude Security and Codex Security at equal compute, with lower precision.
#ai-code-scanning#sast#llm In your stack -
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-2026-75612: yt-mcp add_attachment allows arbitrary local file read before 1.18.1
yt-mcp is a YouTrack MCP server for Claude Code.
PatchUpgrade yt-mcp to version 1.18.1 or later.
-
When Shuffling Fails: Token Recovery from Obfuscated KV Caches
Researchers show shuffled obfuscation of LLM KV caches is weaker than claimed, since repeated tokens reduce the effective state space and allow token recovery.
ReviewReview KV cache sharing and obfuscation assumptions in shared LLM serving infrastructure
Reviewed 2026-11-09 · tools and news update automatically