Data exfiltration through agents

Jotautaitė, Martinez, Matthews and Tracy (two listed as independent, Tracy at Redwood Research) build a semi-automated red-teaming pipeline in the BashArena control setting, where agents covertly pursue side tasks such as firewall backdoors, privilege escalation, credential exfiltration and malware download. MonitoringBench's 2,644 attack trajectories lowered a Claude Opus 4.5 monitor's catch rate at 1% false positives from 94.9% on directly elicited attacks to 60.3% on refined ones.

Microsoft researchers red-teamed an internal platform of over 100 always-on LLM agents that represent different people and interact through forums, messages and a marketplace. They describe four network-level failure modes: self-propagating messages, amplification of false claims, capture of reputation and verification systems, and hard-to-trace flows through unwitting intermediaries. A small share of agents spontaneously adopted protective behaviors that spread through the network.

Check Point Research found that a cloned repository's Claude Code configuration could run hooks, start MCP servers before the user approved them, and redirect API traffic so the user's Anthropic API key was sent to an attacker (CVE-2025-59536, CVE-2026-21852). Anthropic fixed the issues between August and December 2025 by deferring execution and API calls until after the trust dialog.

OpenAI introduced Lockdown Mode, an optional setting that deterministically disables or limits capabilities an attacker could exploit through prompt injection, such as live web access, image support in responses, Deep Research, Agent Mode, live connectors and file downloads. Elevated Risk labels flag network-related features in ChatGPT, Atlas and Codex that carry extra risk. Lockdown Mode first launched for enterprise-type plans, and a June 4, 2026 update says it is rolling out to personal and self-serve Business accounts.

VirusTotal analyzed more than 3,016 OpenClaw skill packages and reports hundreds with malicious behavior, including data exfiltration, backdoors, malware droppers such as Atomic Stealer, and persistent instruction files that manipulate the agent. One publisher accounted for 314 malicious skills; VirusTotal added native scanning of skill packages.

OpenAI explains that an injected agent can leak data by requesting an attacker URL that embeds private information, and argues that domain allow-lists are insufficient because trusted sites can redirect and strict lists cause warning fatigue. Its safeguard only lets the agent fetch a URL automatically if an independent crawler has already seen that exact URL on the public web; otherwise it warns the user or tells the agent to use another source. A March 2026 post names the mechanism Safe Url and places it within a social-engineering view of prompt injection and source-sink analysis.

Miggo Security reports that instructions in a calendar event description stayed dormant until the user asked Gemini about their schedule, then led Gemini to summarize the user's private meetings into a new event the attacker could view. Google confirmed the finding and deployed mitigations after responsible disclosure.

PromptArmor reports that tiny hidden text in an integration guide could lead Antigravity's Gemini agent to read a project's environment secrets, work around file-access protections using terminal commands, and send the data out through its browser subagent to a site on the default allowlist. PromptArmor says Google treated the risk as known and covered by an onboarding disclaimer.

AppOmni reports that instructions planted in an ordinary ServiceNow record could cause a low-privilege Now Assist agent to discover and task a more privileged agent, leading to record changes, data access and email exfiltration. The behavior follows default settings that group agents into teams and make them discoverable; ServiceNow called it intended and updated its documentation.

Anthropic describes sandboxing for Claude Code that restricts file writes to permitted directories and routes network traffic through a proxy that only allows approved hosts, so a prompt-injected agent cannot modify sensitive files or exfiltrate data freely. Anthropic says internal use showed an 84% reduction in permission prompts, and it released the sandbox runtime, built on bubblewrap and macOS seatbelt, as an open-source research preview.

Legit Security found that instructions in hidden pull request comments were processed by Copilot Chat for any user viewing the PR, and that GitHub's Camo image proxy could be used to encode private repository content into a sequence of image requests that bypassed the content security policy. Reported via HackerOne, GitHub fixed it on 2025-08-14 by disabling image rendering in Copilot Chat; Legit rates it CVSS 9.6.

Google DeepMind introduced CodeMender, an agent built on Gemini Deep Think models that combines static and dynamic analysis, fuzzing, differential testing and SMT solvers with LLM-based critique to generate and validate security patches. DeepMind reports 72 security fixes upstreamed to open-source projects over six months, all reviewed by human researchers before submission; in May 2026 Google said it would fold CodeMender into its enterprise agent platform.

MITRE ATLAS version 5.0.0 added a set of techniques for attacks on AI agents, including agent context poisoning of memory and threads, modifying agent configuration, credential theft from agent configuration, and exfiltration via agent tool invocation, and renamed LLM Plugin Compromise to AI Agent Tool Invocation. Version 5.1.0 (November 6, 2025) added agent-specific mitigations such as tool permission configuration and human-in-the-loop for agent actions.

Zenity Labs presented at Black Hat USA 2025 a set of zero- and one-click prompt injection chains, including a shared document causing ChatGPT Connectors to search a victim's Google Drive for API keys and leak them through image rendering, and a poisoned email steering a Copilot Studio agent to disclose CRM data. CSO Online reports that OpenAI and Microsoft deployed fixes for the specific demonstrated techniques.

SafeBreach researchers showed that instructions in calendar invite titles or email subjects, retrieved when a user asked Gemini about their schedule, could trigger tool misuse such as deleting events, opening URLs, exfiltrating email subjects, and controlling connected smart-home devices. They disclosed to Google in February 2025, and Google deployed layered mitigations including confirmations, URL handling policies and injection classifiers.

The European Commission received the final General-Purpose AI Code of Practice, whose Safety and Security chapter applies to providers of models with systemic risk under Article 55 of the AI Act. The chapter treats cyber offence as one of four specified systemic risks, requires a security goal covering non-state external and insider threats, and sets serious incident reporting deadlines that include five days for serious cybersecurity breaches.

Simon Willison argues that an agent becomes exploitable for data theft when it combines access to private data, exposure to untrusted content, and the ability to communicate externally. He advises users to avoid combining all three, points developers to design-pattern mitigations, and argues that guardrails catching most attacks are inadequate in a security setting.

Aim Labs disclosed a zero-click chain in which an email containing hidden instructions, once retrieved by Microsoft 365 Copilot, could cause Copilot to embed internal data in an auto-loaded image request to an attacker. Microsoft rated CVE-2025-32711 critical, fixed it server-side in May 2025, and stated there was no evidence of real-world exploitation.

Invariant Labs demonstrated that a malicious issue in a public repository could lead an agent using the GitHub MCP server to read the user's private repositories and publish the data in a public pull request. The firm tested with Claude 4 Opus and argues there is no server-side patch because the flaw lies in agent permissions, recommending per-session repository scoping and runtime monitoring.

Anthropic activated ASL-3 protections for Claude Opus 4 as a precaution because it could not rule out ASL-3 CBRN risk; the announcement does not cite cyber capability as the trigger. The ASL-3 security standard it describes includes more than 100 controls to protect weights, two-party authorization for weight access, and egress bandwidth controls against exfiltration.

Legit Security reports that hidden instructions in merge requests, comments or code could steer GitLab Duo, combined with unsanitized HTML in streamed responses, to leak private project code and confidential issues. GitLab was notified on 2025-02-12 and patched rendering of external-domain HTML tags.

Invariant Labs describes tool poisoning, in which instructions hidden in an MCP tool's description are visible to the model but not to the user, and shows proof-of-concept exfiltration of local files through an MCP client. It also describes rug pulls, where a server changes tool descriptions after approval, and shadowing, where one server's descriptions alter how the agent uses another server's tools. Recommended mitigations include showing full tool descriptions, pinning tool versions with checksums, and cross-server isolation.

NIST released the 2025 edition of its adversarial machine learning taxonomy, co-authored with the UK AI Security Institute and US AI Safety Institute staff. Unlike the 2023 edition, it includes a section on the security of agents, noting that tool-using agents are exposed to direct and indirect prompt injection and that hijacking can lead to arbitrary code execution or data exfiltration.

Debenedetti and colleagues (Google, Google DeepMind, ETH Zurich) propose CaMeL, which extracts control flow from the trusted user query so untrusted data cannot change which actions run, and attaches capabilities to data to block unauthorized flows. On AgentDojo the first version reported 67% of tasks solved with provable security; the June 2025 revision, with newer models, reports 77% versus 84% for an undefended system.