Prompt injection is untrusted text that an AI model mistakes for an instruction. An email can influence an AI agent when the agent reads or retrieves that message and includes its contents in the model’s context. The email does not grant itself new authority: what the agent can disclose or do depends on the data and tools it already has access to, and on the safeguards around them.
What is prompt injection?
Prompt injection happens when attacker-controlled text is interpreted by a language model as an instruction rather than as material to analyze. The text can be visible or concealed with formatting or non-printing characters, but it does not need to be hidden or placed in a special kind of attachment. Microsoft notes that even plain text can carry an indirect prompt injection.
The distinction between two common forms is where the malicious instruction comes from:
- Direct prompt injection: the person using the AI enters the instruction directly.
- Indirect prompt injection: the instruction arrives inside content the AI is asked to process, such as an email, document, web page, or tool response.
The underlying problem is that the model receives the hostile text alongside the legitimate request and may confuse data it should process with instructions it should follow. Microsoft’s July 2025 security research article calls indirect prompt injection “an inherent risk that arises from the probabilistic language modelling, stochastic generation, and linguistic flexibility of modern LLMs.” The same article said indirect prompt injection was the top entry in OWASP’s 2025 Top 10 for LLM Applications and Generative AI; that is a dated ranking, not a permanent one. Read Microsoft’s explanation of indirect prompt injection.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute#1 Best Overall
- POWERFUL SECURITY KEY: The Security Key C NFC is the essential physical passkey for protecting your digital life from phishing attacks. It ensures only you can access your accounts.
- WORKS WITH 1000+ ACCOUNTS: Compatible with Google, Microsoft, and Apple. A single Security Key C NFC secures 100 of your favorite accounts, including email, password managers, and more.
- FAST & CONVENIENT LOGIN: Plug in your Security Key C NFC via USB-C and tap it, or tap it against your phone (NFC) to authenticate. No batteries, no internet connection, and no extra fees required.
- TRUSTED PASSKEY TECHNOLOGY: Uses the latest passkey standards (FIDO2/WebAuthn & FIDO U2F) but does not support One-Time Passwords. For complex needs, check out the YubiKey 5 Series.
- BUILT TO LAST: Made from tough, waterproof, and crush-resistant materials. Manufactured in Sweden and programmed in the USA with the highest security standards.
How can one email influence an AI agent?
The chain is straightforward: an attacker controls or influences email content, the assistant reads that content, and the text enters the model’s context. The model may then treat the text as a command while carrying out the user’s task. For example, a request to summarize a message can expose the model to instructions embedded in that message, even though the user did not ask the agent to obey them.
- An attacker places instruction-like text in a message sent to a user or otherwise available to the agent.
- The agent retrieves or reads the email to summarize it or complete another task.
- The message becomes part of the context the model uses to decide what to say or do.
- The model may follow the embedded instruction, subject to the application’s permissions and other controls.
“Control” here means influence over the model’s behavior or an attempt to steer its next step. It does not mean that the email can bypass every safeguard or grant new permissions. If an agent cannot access a particular data source or perform a particular action, the email alone cannot give it that access. If it can access sensitive information or use tools such as sending email, however, a successful injection may steer it toward disclosure or unintended actions already within its capabilities.
What could go wrong—and what does not count as a compromise?
Possible consequences include a manipulated answer, disclosure of information available to the agent, or an action the user did not request. Microsoft describes attempts to extract user data and, in an email-capable application, send deceptive messages. The risk therefore depends on the agent’s connected data, tool permissions, and independent controls—not just on the wording in the email.
Rank #2
- POWERFUL SECURITY KEY: The YubiKey 5 NFC is the most versatile physical passkey, protecting your digital life from phishing attacks. It ensures only you can access your accounts
- WORKS WITH 1000+ ACCOUNTS: Compatible with popular accounts like Google, Microsoft, and Apple. A single YubiKey 5 NFC secures 100+ of your favorite accounts, including email, password managers, and more
- FAST & CONVENIENT LOGIN: Plug in your YubiKey 5 NFC via USB and tap it, or tap it against your phone (NFC), to authenticate. No batteries, no internet connection, and no extra fees required
- MOST SECURE PASSKEY: Supports FIDO2/WebAuthn, FIDO U2F, Yubico OTP, OATH-TOTP/HOTP, Smart card (PIV), and OpenPGP. That means it’s versatile, working almost anywhere you need it
- PRIMARY & SPARE KEYS: Just like having a spare house key, we recommend buying two YubiKeys - one for daily use and one as a spare. That way you’ll never get locked out of your accounts
A model being influenced is not automatically a security vulnerability. Microsoft distinguishes influence from security impact: it becomes a vulnerability when the resulting behavior causes harm, such as data exfiltration or an unintended action. An attempted injection that is ignored or contained is not the same thing as a successful compromise.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →What Microsoft’s email-agent challenge does—and does not—show
Microsoft’s LLMail-Inject challenge ran from December 2024 through February 2025. It simulated an LLM-connected email client that could read messages and act on a user’s behalf, including sending email. Participants tried to cause an action the user had not requested while bypassing the challenge’s defenses. Microsoft describes the LLMail-Inject challenge.
Microsoft reported 621 registered participants grouped into 224 teams, with 370,724 submissions. These are challenge participation and submission counts—not real-world attack counts, a success rate, or evidence of how often email agents are compromised. The available evidence does not establish a reliable prevalence or success-rate figure for real-world email prompt-injection attacks.
Rank #3
- Security Key : Protect your online accounts against unauthorized access by using FIDO2 and U2F authentication with T110. It's the world's most protective security key that works with windows, Mac OS, Linux as well as Chrome, Firefox, Edge and many other major browsers.
- Certified with the new FIDO2 standard, T110 provides the benefit of fast login and strong protection against phishing, account takeover as well as many other online attactks.
- Works with : Bank of America, Github, Google, Microsoft, DUO, Twitter, Facebook, Dropbox, Apple, ebay, BINANCE, mor and more.
- Fits USB-A port : Insert the T110 security key into the USB-A port of each service and log in conveniently with one touch
- For the driver download and user guide, please visit TrustKey Solutions Home support page.
How to reduce the risk
No single content filter can guarantee that an agent will resist prompt injection. Microsoft describes detection as one layer in a defense-in-depth approach and characterizes prompt injection as an inherent risk of modern LLMs. Stronger designs also limit what an agent can access and what it can do if it follows hostile text.
Inspect email at ingress, but understand the scope
Microsoft Defender for Office 365 documents scanning inbound messages and assessing subject and body content, including HTML, hidden or off-screen text, quoted or forwarded text, and normalized encoded segments. Its documented focus includes attempts to exfiltrate data through URLs, reveal system prompts, or discover tools. Microsoft also says ordinary instruction-like wording cannot simply be blocked without disrupting legitimate business messages, so runtime safeguards remain important. These are documented capabilities and limits for Microsoft’s product, not a guarantee about all email filters. See Microsoft’s documentation on indirect prompt-injection protection in email.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Separate untrusted content from trusted instructions
At runtime and during retrieval, systems can identify or classify untrusted content and limit how it affects the agent’s instructions. Microsoft describes Prompt Shields as a probabilistic classifier and notes that defenses may be evaded. Treat detection as a risk-reduction measure, not proof that an input is safe.
Rank #4
- POWERFUL SECURITY KEY: The YubiKey 5C NFC is the most versatile physical passkey, protecting your digital life from phishing attacks. It ensures only you can access your accounts
- WORKS WITH 1000+ ACCOUNTS: Compatible with popular accounts like Google, Microsoft, and Apple. A single YubiKey 5C NFC secures 100+ of your favorite accounts, including email, password managers, and more
- FAST & CONVENIENT LOGIN: Plug in your YubiKey 5C NFC via USB and tap it, or tap it against your phone (NFC), to authenticate. No batteries, no internet connection, and no extra fees required
- MOST SECURE PASSKEY: Supports FIDO2/WebAuthn, FIDO U2F, Yubico OTP, OATH-TOTP/HOTP, Smart card (PIV), and OpenPGP. That means it’s versatile, working almost anywhere you need it
- PRIMARY & SPARE KEYS: Just like having a spare house key, we recommend buying two YubiKeys - one for daily use and one as a spare. That way you’ll never get locked out of your accounts
Restrict data and tools to the task
Give the agent access only to the information and actions it needs. Fine-grained data controls and tightly scoped tools limit the damage if the model follows an injected instruction. A summarization task, for example, should not automatically require permission to send messages or access unrelated sensitive records.
Put consequential actions behind controls
Block known harmful effects where possible, such as routes that could exfiltrate data. When an action has external or lasting consequences and the risk cannot otherwise be handled adequately, require explicit user approval. Microsoft cites Outlook’s “Draft with Copilot” flow, in which the user approves and sends generated text, as an example of a human-controlled step.
Log activity and plan for response
Logging, monitoring, detection, and response help teams investigate attempted attacks and possible bypasses. OWASP also identifies related agent risks—including tool abuse, data exfiltration, memory poisoning, goal hijacking, and excessive autonomy—so monitoring should account for more than suspicious text in email. See OWASP’s guidance on agentic AI threats and mitigations. The UK National Cyber Security Centre likewise describes indirect prompt injection through reference content and manipulation through tool responses or connected systems. Read the NCSC’s explanation of prompt injection.
What to check when evaluating an AI email agent
For a practical assessment, ask how the system handles the whole path from incoming content to possible action:
- Coverage: Does protection inspect email at ingress, at model runtime, during retrieval, or at more than one of those points?
- Content visibility: Which formats and content types are examined, including HTML, hidden or off-screen text, quoted material, and encoded segments?
- Permissions: Can the agent access only the data and tools needed for the task, or does it have broad access?
- Action controls: Are known exfiltration paths blocked, and do consequential actions require explicit approval?
- Operational evidence: Can the organization log, investigate, and respond to attempted injections and suspected bypasses?
The cited guidance explains useful control categories, but it does not provide a neutral vendor comparison or comparative effectiveness rates. Treat product claims according to their documented scope and evaluate the permissions and consequence controls around the model as well as its filters.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
