Talk to an Expert

Interactive research

Follow the prompt.
Find the boundary.

Explore five domains of the attack surface, from discovery and delivery to impact and defensive controls.

An analyst examines conflicting instructions reflected across a transparent screen.

Source edition / September 2026

The prompt injection attack surface.

Explore the original map. Select a node for its definition, MITRE ATLAS mapping, documented examples, and source links.

Browse the accessible text version
Reconnaissance & Discovery

How attackers find and profile AI agents before injecting anything. MITRE ATLAS formalized this phase in 2026 with a set of agent-specific reconnaissance techniques — the pre-attack surface most defenders don't yet monitor.

Agent Discovery

Locating deployed agents that are reachable at all.

Enumerate agents behind predictable URLs

Agents hosted on platforms with URL structures derived from environment, tenant, or resource IDs can be enumerated by guessing the pattern — no credentials needed to find them.

From the source edition: MITRE ATLAS AML.T0006.000, contributed from Zenity Labs research (2026).

Source mapping: AML.T0006.000 · Enumerate Hosted AI Resources · Reconnaissance

Search open web for embedded agents

Search results and backlink analysis surface embedded agentic interfaces and let an attacker reverse-engineer internal identifiers from public pages.

From the source edition: MITRE ATLAS AML.T0006 (expanded), 2026.

Source mapping: AML.T0006 · Active Scanning · Reconnaissance

Scan internet databases (Shodan/Censys)

Third-party internet-wide scan services are queried to discover exposed AI services and agent endpoints at scale.

From the source edition: MITRE ATLAS AML.T0000.003, 2026.

Source mapping: AML.T0000.003 · Scan Databases · Reconnaissance

Infrastructure Profiling

Mapping the platform an agent runs on.

Query provider metadata APIs

Provider metadata endpoints are interrogated to enumerate available AI agents and infrastructure details.

From the source edition: MITRE ATLAS AML.T0006.001, 2026.

Source mapping: AML.T0006.001 · Query Platform Metadata APIs · Reconnaissance

Scan for exposed AI infrastructure

Attackers probe for misconfigured or publicly reachable AI systems and their supporting infrastructure.

From the source edition: MITRE ATLAS AML.T0006.002, 2026.

Source mapping: AML.T0006.002 · Scan for Exposed AI Infrastructure · Reconnaissance

Misconfigured / publicly exposed services

Missing authentication and permissive defaults hand unauthenticated attackers a working entry point into an agent.

From the source edition: MITRE ATLAS AML.T0132, 2026.

Source mapping: AML.T0132 · Misconfigured or Publicly Exposed AI Services · Initial Access

Capability Enumeration

Learning what a reachable agent can actually do.

Probe public trigger channels

Webhooks, agent email addresses, and other public trigger channels are poked to see what responds and how — mapping the ways an agent can be invoked.

From the source edition: MITRE ATLAS AML.T0006.003, 2026.

Source mapping: AML.T0006.003 · Probe AI Agent Trigger Channels · Reconnaissance

Discover agent runtime capabilities

Direct interaction with an agent enumerates its available tools and permission scopes, so the attacker knows exactly what an injection could reach.

From the source edition: MITRE ATLAS AML.T0133, 2026.

Source mapping: AML.T0133 · Discover AI Agent Runtime Capabilities · Discovery

Injection Vectors

How attacker instructions reach the model. The fastest-moving domain — most 2026 additions are new places instructions can live, not new ways to phrase them.

Direct Input

The attacker interacts with the model directly as a user.

Attacker-as-user prompting

Malicious instructions typed straight into the chat or API — the classic jailbreak surface. Low sophistication, but still the front door for testing what a deployed model will do.

Source mapping: AML.T0051.000 · LLM Prompt Injection: Direct · Execution

Malicious file attachments

Uploaded documents carry instructions in white-on-white text, metadata fields, or embedded objects that the model reads even though a human reviewer would not.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

User-Mediated Delivery

A legitimate user unknowingly carries the payload in.

Copy-paste smuggling

Content copied from a website or document brings invisible Unicode or hidden text along with it; the user pastes instructions they cannot see into their own prompt.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Poisoned prompt templates

Shared prompt libraries and “mega-prompt” templates ship with embedded directives that execute for every user who adopts them.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Compromised browser extensions

An extension with page access injects instructions into the context an AI browser or side-panel assistant reads.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Crafted AI assistant links

A shared link opens the victim’s assistant pre-loaded with attacker instructions passed as URL parameters — and can poison cross-session memory so the payload outlives the click.

From the source edition: MITRE ATLAS AML.T0131, contributed from Zenity Labs research (2026).

Source mapping: AML.T0131 · Crafted AI Assistant Links · Initial Access

Retrieved Content

The model fetches the payload itself — email, web, knowledge bases, telemetry.

Web page injection

Hidden instructions in HTML attributes, CSS-suppressed elements, or off-screen text on pages an agent browses or summarizes. Now operational, not theoretical.

From the source edition: Unit 42 documented 22 distinct techniques on live malicious sites, with visible-plaintext concealment the single most common delivery method.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

AI-targeted cloaking

Sites serve different content to AI crawlers and agents than to human visitors, so the injection is invisible to anyone who checks the page manually. Now formalized as MITRE ATLAS AML.T0134.

From the source edition: First observed case: AI ad-review evasion — hidden instructions that let malicious ads pass automated moderation (Dec 2025, published Mar 2026).

Source mapping: AML.T0134 · AI Targeted Cloaking · Defense Evasion

Email & calendar payloads

Instructions in a message body, invite description, or subject line execute when an assistant summarizes the inbox or calendar. The premier zero-click vector.

From the source edition: EchoLeak (CVE-2025-32711) exfiltrated M365 data via one crafted email; Zenity’s 2026 “PleaseFix” chain hijacked five AI browsers the same way.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

RAG & knowledge-base poisoning

Malicious records planted in vector stores, wikis, or indexed documents surface as trusted context whenever retrieval touches them — one poisoned entry can steer many users’ sessions.

Source mapping: AML.T0070 · RAG Poisoning · Persistence

Log & telemetry injection

Attacker-controlled strings in logs, alerts, and dashboards become instructions when an AI assistant analyzes them — turning SOC and observability tooling into an injection surface.

From the source edition: “GrafanaGhost” (Apr 2026): poisoned Grafana log entries drove an AI assistant to exfiltrate data through Markdown rendering.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Browser history poisoning

Planting fake entries in browsing history that agentic browsers later treat as trusted facts about the user and their intent.

From the source edition: Zenity’s “HistoryFixing” technique, demonstrated at Black Hat USA 2026.

Source mapping: AML.T0080 · AI Agent Context Poisoning · Persistence

Tooling & Supply Chain

The agent’s own tooling metadata carries the payload.

Tool-description poisoning

Instructions hidden in an MCP server’s tool descriptions execute when the agent loads the tool — no user-visible content involved. Now an OWASP-catalogued attack.

Source mapping: AML.T0110.000 · AI Agent Tool Poisoning: Definition and Instructions · Persistence

Rug-pull tool updates

A tool behaves cleanly at review time, then a later update swaps in malicious descriptions or behavior after trust is established.

Source mapping: AML.T0109 · AI Supply Chain Rug Pull ·

Weaponized skill & rules files

Per-project config files — rules files, skills, agent instructions — as a malware-style delivery channel, complete with supply-chain distribution.

From the source edition: Zenity’s “Promptware” research at Black Hat 2026 treated skill files, MCP servers, and misaligned models as an end-to-end malicious supply chain.

Source mapping: AML.T0081 · Modify AI Agent Configuration · Persistence

Poisoned repo metadata & CI

Commit messages, issues, branch names, and PR text execute when coding agents and CI-integrated assistants process them.

From the source edition: Three CVEs in Anthropic’s MCP Git server (Jan 2026) via commit messages and branch names; booby-trapped GitHub issues steering CI agents.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Synthetic tool outputs

Forged tool results injected into the agent loop — the model trusts a “tool response” that no tool produced.

Source mapping: AML.T0110.002 · AI Agent Tool Poisoning: Runtime Response · Persistence

Agent Ecosystem

Payloads that move between agents, sessions, and memories.

Agent-to-agent injection

In multi-agent systems, one compromised agent’s output becomes another’s instructions — injections propagate across the mesh.

Source mapping: AML.T0118 · Autonomous AI Agent Communication · AI Attack Adaptation

Persistent memory poisoning

Instructions written into long-term memory features re-execute in every future session, long after the original poisoned content is gone.

Source mapping: AML.T0080.000 · AI Agent Context Poisoning: Memory · Persistence

Cross-session prompt residue

Leftover context — cached reasoning, session artifacts — reused across sessions and users becomes an unintended carrier for stale attacker instructions.

From the source edition: Demonstrated against GitHub, Slack, and Jira agent integrations at Black Hat 2026.

Source mapping: AML.T0080.001 · AI Agent Context Poisoning: Thread · Persistence

Prior-model-output injection

Output from one model, stored and later fed to another, carries instructions across trust boundaries neither system checks.

Source mapping: AML.T0051.001 · LLM Prompt Injection: Indirect · Execution

Multimodal Carriers

Non-text channels that smuggle instructions past text-based defenses.

Image-embedded instructions

Near-invisible text overlaid on images defeats content labeling and scanner-based defenses; the vision model reads what the human eye misses.

From the source edition: Brave’s Black Hat 2026 research bypassed layered defenses — strong prompts, content labeling, tool scanning, verifier models — with image overlays.

Source mapping: AML.T0129 · Triggers in Multimodal Inputs · Defense Evasion

Audio & transcription poisoning

Adversarial audio yields transcripts containing instructions the speaker never audibly said.

Source mapping: AML.T0129 · Triggers in Multimodal Inputs · Defense Evasion

UI & markup camouflage

Instructions tucked into spoiler tags, alt text, and ARIA attributes — read by agents, hidden from people.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

File & media metadata triggers

Instructions hidden in the metadata of image, audio, video, or document files — parsed by the model even though no reviewer opens the metadata.

From the source edition: MITRE ATLAS AML.T0129 (Triggers in Multimodal Inputs), 2026.

Source mapping: AML.T0129 · Triggers in Multimodal Inputs · Defense Evasion

Manipulation Techniques

How a delivered payload subverts model behavior — the domain jailbreak research has mapped deepest.

Instruction Override

Directly rewriting what the model believes its rules are.

Rule addition & nullification

“New rule:” / “ignore previous instructions” — appending, negating, or substituting operating rules.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Authority & system spoofing

Payloads impersonate system messages, administrators, or claim pre-authorization the user never gave.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Refusal suppression

Forbidding refusal language itself — apologies, warnings, “I can’t” — so declining stops being an available output.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

Dormant trigger rules

Sleeper instructions planted early that stay inert until a trigger phrase activates them turns later — evading point-in-time inspection.

From the source edition: Named “Trigger-Activated Rule Addition” in CrowdStrike’s July 2026 taxonomy update.

Source mapping: AML.T0051.002 · LLM Prompt Injection: Triggered · Execution

Obfuscation & Encoding

Making the payload unreadable to filters but readable to the model.

Invisible Unicode smuggling

Zero-width and Unicode tag characters carry instructions no human reviewer can see in rendered text.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Encoded payloads

Base64, ciphers, and attacker-invented encodings the model helpfully decodes — sometimes assembled at runtime from fragments.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Payload splitting & reassembly

Instructions fragmented into pieces that look benign individually and only become malicious when the model recombines them.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Multilingual & low-resource evasion

Payloads in languages or dialects where safety training is thinnest, or switched mid-prompt to shake classifiers.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Homoglyph & typographic substitution

Look-alike characters, deliberate typos, and formatting tricks that break keyword matching while preserving meaning.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Boundary Exploitation

Attacking the seams between system, user, and data.

Special-token & delimiter mimicry

Payloads imitate chat-template tokens and formatting boundaries so untrusted content gets parsed as privileged instructions.

From the source edition: Named “Special Token Injection” in CrowdStrike’s July 2026 update.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Forged conversation history

Fake prior turns — including fabricated assistant agreements — imported as few-shot context the model then continues.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Format-boundary confusion

Code fences, quotes, and markup used to make instructions look like inert data, or data look like instructions.

Source mapping: AML.T0068 · LLM Prompt Obfuscation · Defense Evasion

Reasoning-Layer Attacks

Targeting how the model thinks rather than what it is told.

Reasoning-channel injection

Injecting into hidden reasoning and execution channels rather than the visible prompt — hijacking chain-of-thought and internal tool calls directly.

From the source edition: Palo Alto’s Black Hat 2026 “Reasoning Injection Attack” hijacked hidden python execution channels in a full sandbox-compromise chain.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Reasoning hijack & overload

Steering deliberation with contrived logic, or exhausting it with complexity so safety checks get crowded out.

Source mapping: AML.T0051 · LLM Prompt Injection · Execution

Safety-token suppression

Blocking safety-related tokens and reasoning tags so the model cannot produce its own guardrail language.

From the source edition: Named “Cognitive Token Suppression” in CrowdStrike’s July 2026 update.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

Multi-Turn & Persistence

Attacks that unfold across turns and sessions.

Gradual steering

Crescendo-style escalation: each turn moves one plausible step until the destination is somewhere the model would have refused outright.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

In-session protocol setup

Establishing codewords, variables, or private rules early that later turns invoke innocuously.

Source mapping: AML.T0094 · Delay Execution of LLM Instructions · Defense Evasion

Memory implantation

Asking the assistant to “remember” attacker rules so its own memory feature becomes the persistence mechanism.

Source mapping: AML.T0080.000 · AI Agent Context Poisoning: Memory · Persistence

Social & Cognitive Framing

Persuasion aimed at the model’s judgment.

Roleplay & fictional framing

Harmful requests wrapped in personas, stories, or games so compliance reads as staying in character.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

Urgency & authority pressure

Manufactured emergencies, claimed credentials, and false stakes that make refusal feel like the harmful act.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

Detached-reality framing

“Test mode”, “sandboxed environment”, “outputs don’t matter here” — convincing the model consequences don’t apply.

Source mapping: AML.T0054 · LLM Jailbreak · Defense Evasion

Attack Outcomes

What injections achieve once they land — the impact dimension most taxonomies stop short of.

Data Exfiltration

Getting secrets out through the model’s own channels.

Zero-click exfiltration

Data theft with no user interaction: the assistant processes attacker content in the background and leaks on its own.

From the source edition: EchoLeak (CVE-2025-32711) against Microsoft 365 Copilot — one email, no clicks.

Source mapping: AML.T0057 · LLM Data Leakage · Exfiltration

Rendering-channel leaks

Markdown images and crafted links smuggle data into URLs that fire automatically when the response renders.

Source mapping: AML.T0077 · LLM Response Rendering · Exfiltration

Credential & vault theft

Agents with browser or filesystem reach are steered into harvesting stored credentials.

From the source edition: Zenity’s Comet chain exploited the 1Password extension to steal an entire password vault from a spoofed calendar invite.

Source mapping: AML.T0086 · Exfiltration via AI Agent Tool Invocation · Exfiltration

Action Hijacking

The agent’s legitimate capabilities, misdirected.

Fraudulent transactions

Injected instructions drive purchases, transfers, and subscription changes with the victim’s standing payment authority.

From the source edition: CSA documented forced $5,000 PayPal transfer attempts and Stripe subscription fraud among in-the-wild injection intents.

Source mapping: AML.T0053 · AI Agent Tool Invocation · Execution

Outbound phishing & messaging

The victim’s own assistant sends phishing and scam messages from their accounts, inheriting their trust.

From the source edition: ChatGPT Atlas hijack sent WhatsApp phishing from the victim’s identity (Black Hat 2026).

Source mapping: AML.T0053 · AI Agent Tool Invocation · Execution

Destructive operations

Recursive file deletion, dropped databases, deleted cloud infrastructure — sabotage through granted permissions.

From the source edition: Zenity’s demos included Gemini deleting AWS servers and Copilot Edge corrupting SQL databases via localhost access.

Source mapping: AML.T0101 · Data Destruction via AI Agent Tool Invocation · Impact

Privilege Escalation Chains

Injection as the first stage of classic exploitation.

Prompt-to-RCE

Injections chain into deserialization, path traversal, and command execution bugs in agent frameworks — a prompt becomes an initial-access vector.

From the source edition: Microsoft’s May 2026 research and Check Point’s 11 framework vulnerabilities across LangChain, CrewAI, and AutoGen.

Source mapping: AML.T0050 · Command and Scripting Interpreter · Execution

Sandbox escape

Breaking out of the model’s execution sandbox into the hosting platform, including cross-tenant reach.

From the source edition: Demonstrated at Black Hat 2026 against ChatGPT’s sandbox and Cloudflare Workers code execution.

Source mapping: AML.T0105 · Escape to Host · Privilege Escalation

Internal network reach

Agents with localhost or intranet access become an SSRF proxy into environments the attacker could never touch directly.

Source mapping: AML.T0053 · AI Agent Tool Invocation · Lateral Movement

Persistence & Lateral Movement

Staying in, and spreading.

Memory persistence

Implanted instructions survive session resets and re-execute indefinitely from the assistant’s own memory.

Source mapping: AML.T0080.000 · AI Agent Context Poisoning: Memory · Persistence

Cross-agent propagation

A payload hops between agents in a workflow, compromising systems the initial injection never touched.

Source mapping: AML.T0061 · LLM Prompt Self-Replication · Persistence

Account takeover chains

One injection pivots across connected services — email, chat, code hosting — into full account compromise.

From the source edition: Zenity’s single-email chain reached Gmail, Drive, Slack, and X account takeover.

Source mapping: AML.T0053 · AI Agent Tool Invocation · Lateral Movement

Kinetic Impact

Injections against agents that control physical systems — data manipulation becomes physical-world compromise. Newly named as an attack class at Black Hat 2026.

Source mapping: · No direct ATLAS technique · Impact

Integrity & Influence

Corrupting decisions rather than stealing data.

Moderation & review evasion

Injected instructions make AI reviewers approve content they should reject — the first injection class observed operating at scale in the wild.

From the source edition: AI ad-review evasion observed by Unit 42, December 2025.

Source mapping: AML.T0130 · AI Agent Response Biasing · Impact

Decision manipulation

Skewing AI-assisted screening, ranking, and recommendations — résumé filtering, vendor scoring, triage queues.

From the source edition: Recruitment-screening manipulation documented among in-the-wild injection intents.

Source mapping: AML.T0130 · AI Agent Response Biasing · Impact

Defensive Controls

No single control stops prompt injection. These layers assume injection will land and limit what it can do.

Architecture

Design decisions that cap the blast radius.

Least-privilege agents

Scope every agent to the minimum tools, data, and identities its task needs — the injection inherits only what the agent holds.

Human-in-the-loop gates

Consequential actions — payments, sends, deletes, publishes — require a person to confirm, outside the model’s control.

Context isolation & trust labeling

Keep untrusted content structurally separate from instructions and track its provenance through the pipeline, so “data” can’t silently become “commands”.

Input & Output Controls

Filtering on the way in, containment on the way out.

Injection scanning & classifiers

Detect known payload patterns in retrieved content. Necessary but insufficient — Black Hat 2026 research bypassed layered scanner defenses with image overlays alone.

Egress & rendering controls

Block auto-loading images, restrict outbound URLs, and constrain markdown rendering — closing the channels exfiltration actually uses.

Tool allowlisting & scoped credentials

Agents call only vetted tools with short-lived, narrowly scoped credentials; an injected instruction can’t invoke what isn’t there.

Supply Chain Hygiene

Treating agent tooling like the software supply chain it is.

MCP server & tool vetting

Review and pin tool versions, hash descriptions, and re-verify on every update — rug-pull attacks depend on unwatched changes.

Skill & rules file review

Config files that steer agents get code-review treatment, including scanning for invisible Unicode.

Detection & Response

Assume compromise; watch behavior.

Agent action telemetry

Log every tool call, retrieval, and output channel; alert on anomalies like unexpected egress or privilege use.

Continuous red teaming

Test deployed agents with current techniques — this taxonomy — on a schedule, not once at launch.

Governance

Making the risk owned, not assumed.

OWASP LLM Top 10 alignment

Map controls to LLM01 and its siblings so coverage is auditable and gaps are explicit.

AI-feature vendor evaluation

Procurement asks how a product isolates untrusted content, gates actions, and handled its last injection disclosure.

Original Defensive visualization, mapped to MITRE ATLAS. References, examples, and mappings retain the September 2026 edition’s context; this is an educational aid, not a complete security assessment.

From research to controls

Test the boundaries that matter.

Efficacy first. Always.

What’s next
for your world?

Bring us the problem, the constraints, and the outcome you need. We’ll help shape a practical way forward.

Let’s talk about it