Agentic AI’s Attack Surface Is Already Exploited

On 29 September 2025, Koi Security researchers flagged the first confirmed malicious Model Context Protocol server in the wild. A rogue npm package named postmark-mcp — a near-identical copy of Postmark Labs’ official library — had spent two weeks silently BCC’ing every email sent through it to an attacker-controlled address. It accumulated 1,643 downloads before removal. The attacker’s tool: one line of code. The victim’s visibility: zero.

agentic AI security overview graphic

The risks of agentic AI security are no longer theoretical, and security leaders are reworking controls around them. For the broader industry response, see our coverage of Frontier Model Forum distillation security.

That incident is not an outlier. It is one entry in a documented sequence of confirmed attacks across 2025 that collectively define a new category of enterprise exposure — one where agentic AI security challenges cannot be resolved by conventional endpoint detection, DLP, and WAF tooling.

The Attack Surface Agentic Systems Create

Traditional LLM security focused on what a model would say. Agentic AI security concerns what a model will do. An agent equipped with tool-calling capabilities — web browsing, code execution, email dispatch, API access, file system operations — operates with a trust level and an action radius that exceeds most human users on the same network.

Vendors are also building defences into model access programmes, which makes agentic AI security a question of governance as much as engineering. Readers can compare approaches in our analysis of Anthropic’s Glasswing cybersecurity work.

The attack surface has three distinct layers. First, the model itself can be manipulated through its inputs. Second, the tools the model calls can be compromised at the source. Third, the agent’s permissions — the scope of what it can access and act on — are routinely misconfigured far above what any given task requires.

OWASP’s GenAI Security Project, drawing on more than 100 security researchers, published a dedicated Top 10 for Agentic Applications in December 2025, separate from its existing LLM Top 10. The list formalises what practitioners had already encountered: prompt injection (AG01), memory poisoning (AG03), tool and plugin misuse (AG04), and supply chain vulnerabilities (AG09) represent the leading risk categories.

Prompt injection tops both the new Agentic Top 10 and the existing LLM Top 10 — it is, per OWASP’s own framing, the root vulnerability from which most other agentic exploits branch.

Documented Incidents: Prompt Injection Moves to Production

The theoretical risk of prompt injection has been well-documented since 2023. What 2025 added was a confirmed incident record proving that agentic AI security failures have moved directly into production enterprise systems.

CVE-2025-32711, disclosed by Aim Security and dubbed EchoLeak, is the clearest illustration of the severity ceiling. The vulnerability, rated CVSS 9.3, exploited Microsoft 365 Copilot’s document processing to achieve zero-click data exfiltration. An attacker sends an email to the target’s Outlook inbox containing hidden adversarial instructions — no interaction required on the victim’s side.

When Copilot processes the email (to summarise, triage, or analyse it), it follows the embedded instructions, autonomously exfiltrating data from OneDrive, SharePoint, and Teams to attacker-controlled infrastructure.

According to Hack The Box’s published analysis, the attack chained four bypasses: evasion of Microsoft’s XPIA classifier, link redaction bypass via reference-style Markdown, exploitation of auto-fetched images, and abuse of a Teams content security policy allowlist. Microsoft patched server-side in June 2025. No client-side action was required.

CVE-2025-53773, patched in Microsoft’s August 2025 Patch Tuesday, demonstrated a complementary vector through developer tooling. Researchers at Embrace The Red documented the attack chain: malicious instructions embedded in a public GitHub issue or README cause GitHub Copilot’s VS Code agent to write “chat.tools.autoApprove”: true to .vscode/settings.json. That setting disables all user confirmation prompts for Copilot’s tool calls.

With that gate removed, a subsequent conditional injection issues terminal commands — achieving remote code execution on the developer’s machine across Windows, macOS, and Linux. The researchers noted the core architectural flaw: “the edits are immediately persistent, they are not in-memory as a diff to review.”

Palo Alto Networks Unit 42’s December 2025 telemetry on web-based indirect prompt injection documented 12 confirmed real-world attack cases and characterised the payload landscape: 37.8% of observed injections used visible plaintext, 19.8% used HTML attribute cloaking, and 16.9% used CSS rendering suppression. Data destruction attempts represented 14.2% of observed attacker intent — not just data theft, active deletion.

The MCP Supply Chain: An Unpatrolled Dependency Layer

Model Context Protocol servers are the connective tissue between AI agents and the tools they use — file systems, databases, APIs, communication platforms — and a critical focus area for agentic AI security. Anthropic open-sourced the protocol in November 2024; adoption accelerated rapidly through 2025 as developers published MCP servers to npm and other registries with no vetting process comparable to what major package ecosystems apply to general software.

The breach timeline through 2025 is precise. In April 2025, a malicious MCP server exfiltrated WhatsApp chat history through tool poisoning — manipulating tool descriptions to redirect output to attacker-controlled numbers, per authzed.com’s published breach timeline. In May 2025, a prompt injection via GitHub MCP’s public issues caused private repository contents to leak into public pull requests due to overly broad Personal Access Token scopes.

In June 2025, Anthropic’s own MCP Inspector developer tool contained an unauthenticated remote code execution flaw, exposing filesystems, API keys, and environment secrets on developer machines.

July 2025 brought CVE-2025-6514 in the mcp-remote package — a CVSS 9.6 OS command injection vulnerability affecting 437,000 installations. In September 2025, the postmark-mcp backdoor demonstrated supply chain compromise at the package level. October 2025 added the Smithery path-traversal incident: a single build configuration flaw exposed a Fly.io API token granting control over 3,000+ MCP server deployments, capturing inbound traffic and downstream API secrets.

The structural problem is that MCP servers operate with the trust level of the agent that invokes them — which is often significantly elevated. A compromised MCP server can request and receive permissions the legitimate application never needed. Unlike SaaS integrations, which are reviewed by security and legal teams, MCP servers are frequently installed by individual developers and immediately granted access to production credentials.

Privilege Escalation and the Confused Deputy Problem

Beyond supply chain, the configuration failures in agentic deployments create what security literature calls the “confused deputy” problem, representing a core agentic AI security vulnerability where a less-privileged party manipulates a more-privileged intermediary into acting on its behalf.

The mechanics in agentic systems are direct. An agent given access to an email system, a database, and a file server for legitimate productivity tasks holds more combined access than any of the individual employees it serves. When an attacker manipulates that agent via an indirect prompt injection — a poisoned document it is asked to summarise, a malicious website it browses during a research task — the attacker inherits the agent’s full permission set, not their own.

IBM’s 2025 Cost of a Data Breach Report quantified the outcome: shadow AI breaches — those involving AI agents operating outside formal governance — averaged $4.63 million per incident, $670,000 more than a conventional breach, per Reco AI’s 2025 year-in-review analysis. The cost differential reflects the difficulty of detection: conventional DLP and endpoint tools have no visibility into what an AI agent decided to do with the data it accessed.

NIST’s updated AI 100-2 E2025 (Adversarial Machine Learning), published in 2025, is the first edition of that taxonomy to explicitly address agentic systems — the 2023 version did not. It covers prompt injection, RAG-targeting attacks, and supply chain risk as distinct categories, per NIST’s publication record. That addition signals that federal guidance is catching up to what attackers already demonstrated in the wild.

What Enterprise Security Teams Need to Do Now

Current controls are insufficient because they were designed for human actors. The required response operates at three levels.

Least privilege for agents, not just users. Within any robust agentic AI security posture, agents must be scoped to precisely the permissions their current task requires, not the permissions they might ever need. This is harder than user RBAC because agents are often deployed with long-lived credentials to avoid re-authentication overhead. NVIDIA’s published sandboxing guidance for agentic workflows recommends OS-level isolation — macOS Seatbelt, Linux seccomp/namespaces — network egress whitelisting, and blocking filesystem writes outside defined workspaces.

These are not AI-specific controls; they are the same principles applied to containerised workloads, applied one layer deeper.

MCP server governance as a supply chain problem. Every MCP server in use is a dependency that requires the same treatment as an open-source library: inventoried, version-locked, signature-verified, and monitored for upstream changes. The Databricks AI Security Framework (DASF v3.0) added Agentic AI as its 13th component, explicitly covering MCP threats alongside memory and planning risks, per the Databricks blog. SBOMs for MCP servers are not yet standard practice; they should be.

Prompt injection detection at ingestion, not output. Microsoft’s July 2025 MSRC guidance introduced “Spotlighting” — using delimiting, datamarking, or encoding to help LLMs distinguish trusted instructions from untrusted external content — alongside Microsoft Prompt Shields, a classifier that detects injection attempts before they reach the model. Per Microsoft’s published guidance, this is a detection layer, not a prevention guarantee.

Microsoft released the Agent Governance Toolkit on 2 April 2026, addressing all 10 OWASP agentic risks through runtime monitoring, per the Microsoft Open Source blog.

On the compliance side: neither SOC 2 nor ISO 27001 contains control categories specific to AI agents. ISO 42001 (AI Management Systems) is the emerging standard, alongside NIST AI RMF. Gartner projects 60% of organisations will have formalised AI governance programs by 2026, per penligent.ai’s compliance analysis. Security teams should not wait for auditors to mandate agent-specific controls — the 2025 incident record makes the case.

Least-Privilege Guardrails and Execution Boundaries for Agentic AI Security

Establishing comprehensive agentic AI security across enterprise environments requires information security teams to abandon passive perimeter defense models. When autonomous agents are granted tool-execution capabilities and API access tokens, robust agentic AI security demands strict runtime least-privilege guardrails. Securing modern agentic AI security architectures requires deterministic command validation and ephemeral credential scoping before any high-privilege action executes.

Mitigating prompt injection and confused deputy attacks within agentic AI security frameworks requires isolated sandbox execution environments. As enterprises connect autonomous systems to internal databases, agentic AI security protocols must inspect multi-turn conversation traces for privilege escalation signatures. Implementing continuous agentic AI security monitoring ensures that automated tool invocations remain strictly within pre-approved operational boundaries.

Ultimately, the maturity of agentic AI security deployments will determine how rapidly regulated industries embrace fully autonomous business workflows. Security leaders who operationalize defense-in-depth agentic AI security frameworks will protect critical enterprise assets from adversarial exploitation while safely unlocking the productivity of agentic automation.

What to Watch

Human-in-the-loop requirements for high-risk tool calls. The pattern in EchoLeak and CVE-2025-53773 is autonomous execution — no human in the approval chain. Requiring explicit user confirmation before agents send emails, execute code, or modify configuration files is an architectural decision that needs to be made at deployment time, not patched in after an incident.

Registry vetting for MCP servers. The MCP ecosystem currently has no equivalent of npm’s security audit tooling or PyPI’s malware scanning. The September 2025 postmark-mcp incident is the first; it will not be the last. Watch for whether Anthropic, npm, or third-party security vendors introduce signing and provenance requirements for MCP packages.

Regulatory movement on agentic AI. The EU AI Act’s high-risk system categories do not explicitly cover agentic deployments; enforcement guidance for 2026 may change that. NIST’s AI RMF is being adopted by US federal contractors as a procurement requirement. Enterprise security teams negotiating AI vendor contracts in 2026 should require vendors to document their agentic security posture against the OWASP Agentic Top 10 as a baseline.

Agent identity and audit logging. If an AI agent exfiltrates data, current SIEM tooling may not distinguish agent activity from user activity — both appear as API calls under the same credential. Agent-specific identity tokens and structured audit trails for every tool invocation are the foundational observability requirement that most deployments have not yet implemented.

Further Reading

This article was produced with AI assistance and reviewed by the editorial team.

Arjun Mehta

Arjun Mehta

Arjun Mehta covers AI infrastructure, semiconductor manufacturing, and the science-to-commercialization pipeline for Next Waves Insight. His reporting focuses on the technical and economic mechanics behind emerging technology — from chip fabrication constraints to the regulatory and funding pathways that determine which breakthroughs actually reach production. He has covered the technology sector for several years, with a focus on connecting primary research and hardware development to their downstream industry impact.

2 thoughts on “Agentic AI’s Attack Surface Is Already Exploited”

Leave a Comment