āĻĒā§āϰāϝ⧁āĻ•ā§āϤāĻŋāĻ—āϤ āĻ—āĻžāχāĻĄ

Spotlighting and Datamarking Defenses

Spotlighting marks or transforms untrusted text so a model has a cue that it is data to process rather than an instruction to follow.

āĻāχ āύāĻŋāĻ°ā§āĻĻ⧇āĻļāĻŋāĻ•āĻž āĻļ⧁āύ⧁āύāĻāφāχ-āĻŦāĻ°ā§āĻŖāĻŋāϤ ¡ ~3 min

āĻāχ āĻĒ⧃āĻˇā§āĻ āĻžāϝāĻŧ3 āĻŽāĻŋāύāĻŋāϟ āĻĒāĻĄāĻŧāĻž āĻšāϝāĻŧ⧇āϛ⧇
  1. āĻ“āĻ­āĻžāϰāĻ­āĻŋāω
  2. āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ
  3. āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ
  4. The Future of Spotlighting and Datamarking Defenses
  5. āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇
  6. āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ
  7. āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

āĻ“āĻ­āĻžāϰāĻ­āĻŋāω

Research describes delimiting, datamarking, and encoding variants, but results are tied to tested models and attacks; no formatting trick creates a guaranteed security boundary.

āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ

Spotlighting is a family of prompt transformations intended to make the provenance of untrusted content more visible in a model’s text context. A 2024 Microsoft-affiliated research paper describes three variants. Delimiting brackets a document with selected markers. Datamarking inserts a marker throughout the text, in the paper’s example replacing whitespace between words. Encoding transforms the document, for example using base64, and instructs the model to interpret the encoded block as data for the assigned task. Each variant also relies on a system instruction that explains how the marked text should be treated. The paper evaluated these methods on a synthetic indirect-injection corpus and older GPT-family model versions. It reported substantially lower attack success in its experiments, including a decrease from above 50% to below 2% for selected settings. Those numbers describe the paper’s particular tasks, attacks, and models; they are not expected production rates or guarantees for current systems. The paper also found encoding could impair underlying tasks for models less capable of decoding the text. This makes per-model utility evaluation important, not optional. Microsoft Foundry documents its Spotlighting preview as a document-attack control that uses base64 transformation and is supported only for models through the Chat Completions API. The docs say this increases document tokens and can cause long inputs to exceed limits, and note a possible user-visible encoding reference. That specific service behavior differs from a generic hand-built technique. In every case, marking helps signal provenance inside a shared text channel; it does not isolate data at a hardware or permission boundary. Test attacks that know the marker, preserve ordinary task quality, and restrict tool permissions so a successful injection has limited consequences.

āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ

A summarizer places a retrieved document between delimiters and states that instructions inside the passage are content to summarize, not commands.

A pipeline inserts a marker between words in a document, following a datamarking approach, then tests whether normal summarization still works.

A system transforms a passage using an encoding and tells a capable model how to interpret it as reference content, then measures task accuracy and attack success.

A security team pairs data marking with restricted tools and tests for attacks that copy, omit, or imitate the chosen marker.

The Future of Spotlighting and Datamarking Defenses

Provenance cues may become more native to model interfaces, but prompt-level markings still share the same context channel as the text they label. The research and product implementations will evolve, and their effects will vary with model versions and attack designs. Teams should retest task quality and adversarial cases after any change, while retaining independent permission checks and monitoring. Dynamic markers could make some simple spoofing attempts harder, but they do not create a separate trusted channel. Review new attack patterns as retrieval sources change.

āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇

āϕ⧇āύ āĻāϟāĻž āϗ⧁āϰ⧁āĻ¤ā§āĻŦāĻĒā§‚āĻ°ā§āĻŖ

  • āĻ–āϰāϚ āĻ“ āĻŦāĻžāĻœā§‡āϟ. āφāĻ°ā§āĻ•āĻŋāĻŸā§‡āĻ•āϚāĻžāϰ⧇āϰ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤāϗ⧁āϞāĻŋ āĻŦāĻ›āϰ⧇āϰ āĻĒāϰ āĻŦāĻ›āϰ āϧāϰ⧇ āĻ•āĻ°ā§āĻŽāĻ•ā§āώāĻŽāϤāĻž āĻāĻŦāĻ‚ āĻ…āĻĒāĻžāϰ⧇āϟāĻŋāĻ‚ āĻ–āϰāϚ āϚāĻžāϞāĻžāϝāĻŧāĨ¤
  • āϏ⧁āĻ¸ā§āĻĒāĻˇā§āϟ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤ. āĻ•āĻžāϰāĻŋāĻ—āϰāĻŋ āĻļāĻŋāĻ•ā§āώāĻž āĻĻāϞāϗ⧁āϞ⧋āϕ⧇ āϏāĻ āĻŋāĻ• āĻ¸ā§āĻŸā§āϝāĻžāĻ• āĻŦ⧇āϛ⧇ āύāĻŋāϤ⧇ āϏāĻžāĻšāĻžāĻ¯ā§āϝ āĻ•āϰ⧇, āĻļ⧁āϧ⧁ āύāϤ⧁āύāϟāĻŋ āύāϝāĻŧāĨ¤
  • āĻŽāĻžāύ āύāĻŋāϝāĻŧāĻ¨ā§āĻ¤ā§āϰāĻŖ. āĻ­āĻžāϞ⧋ āχāĻžā§āϜāĻŋāύāĻŋāϝāĻŧāĻžāϰāĻŋāĻ‚ āĻĒāĻ›āĻ¨ā§āĻĻ āĻ‰ā§ŽāĻĒāĻžāĻĻāύ⧇ āύāĻŋāĻ°ā§āĻ­āϰāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āϘāϟāύāĻž āĻ•āĻŽāĻŋāϝāĻŧ⧇ āĻĻ⧇āϝāĻŧāĨ¤

āϜāĻ¨ā§āϝ āϏāϤāĻ°ā§āĻ•

  • āĻāĻ•āϟāĻŋ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ• āĻ…āĻĒā§āϟāĻŋāĻŽāĻžāχāϜ āĻ•āϰāĻž āĻŦ⧃āĻšāĻ¤ā§āϤāϰ āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽ āĻĻ⧁āĻ°ā§āĻŦāϞāϤāĻž āφāĻĄāĻŧāĻžāϞ āĻ•āϰāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤
  • āĻ…āĻŦāĻ•āĻžāĻ āĻžāĻŽā§‹ āĻāĻŦāĻ‚ āϰāĻ•ā§āώāĻŖāĻžāĻŦ⧇āĻ•ā§āώāϪ⧇āϰ āĻ–āϰāϚ āĻĒā§āϰāĻžāϝāĻŧāχ āĻ…āĻŦāĻŽā§‚āĻ˛ā§āϝāĻžāϝāĻŧāύ āĻ•āϰāĻž āĻšāϝāĻŧāĨ¤
  • āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽāϗ⧁āϞāĻŋ āφāϰāĻ“ āϜāϟāĻŋāϞ āĻšāĻ“āϝāĻŧāĻžāϰ āϏāĻžāĻĨ⧇ āϏāĻžāĻĨ⧇ āϏ⧁āϰāĻ•ā§āώāĻž āĻāĻŦāĻ‚ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āĻĢāĻžāρāĻ• āĻŦāĻžāĻĄāĻŧāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻĒāĻĻāĻ•ā§āώ⧇āĻĒ

  1. āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ⧇āϰ āφāϗ⧇ āĻŦāĻŋāϞāĻŽā§āĻŦ, āϗ⧁āĻŖāĻŽāĻžāύ āĻāĻŦāĻ‚ āĻ–āϰāĻšā§‡āϰ āϞāĻ•ā§āĻˇā§āϝ āύāĻŋāĻ°ā§āϧāĻžāϰāĻŖ āĻ•āϰ⧁āύāĨ¤
  2. āĻŦāĻžāĻ¸ā§āϤāĻŦāϏāĻŽā§āĻŽāϤ āϞ⧋āĻĄ āĻāĻŦāĻ‚ āĻĄā§‡āϟāĻž āĻ…āĻŦāĻ¸ā§āĻĨāĻžāϰ āĻ…āϧ⧀āύ⧇ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ•āĨ¤
  3. āĻ¤ā§āϰ⧁āϟāĻŋ, āĻĒā§āϰāĻŦāĻžāĻš, āĻāĻŦāĻ‚ āĻŦā§āϝāĻŦāĻšāĻžāϰāĻ•āĻžāϰ⧀āϰ āĻĒā§āϰāĻ­āĻžāĻŦ⧇āϰ āϜāĻ¨ā§āϝ āϝāĻ¨ā§āĻ¤ā§āϰ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāĨ¤
  4. āĻ¸ā§āϕ⧇āϞ āĻ•āϰāĻžāϰ āφāϗ⧇ āϰ⧋āϞāĻŦā§āϝāĻžāĻ• āĻāĻŦāĻ‚ āϘāϟāύāĻžāϰ āĻĒā§āϰāϤāĻŋāĻ•ā§āϰāĻŋāϝāĻŧāĻž āĻĒāĻžāĻĨ āĻĒā§āϰāĻ¸ā§āϤ⧁āϤ āĻ•āϰ⧁āύāĨ¤

āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ

āĻĢā§āϰāĻŋ āύāĻŋāωāϜāϞ⧇āϟāĻžāϰ

āĻĒā§āϰāϤāĻŋāĻĻāĻŋāύ⧇āϰ āĻāφāχ āĻŦā§āϰāĻŋāĻĢāĻŋāĻ‚ āĻĒāĻžāύ

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ āϏāĻ•āĻžāϞ⧇ āϤāĻŋāύāϟāĻŋ āϝāĻžāϚāĻžāχāĻ•ā§ƒāϤ āĻāφāχ āĻ¸ā§āĻŸā§‹āϰāĻŋ, āϏāĻšāϜ āχāĻ‚āϰ⧇āϜāĻŋāϤ⧇ āϞ⧇āĻ–āĻžāĨ¤ āϚāĻŋāϰāĻ•āĻžāϞ⧇āϰ āϜāĻ¨ā§āϝ āĻĢā§āϰāĻŋ, āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχāĨ¤

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ⧇ āĻāĻ•āϟāĻŋ āχāĻŽā§‡āϞāĨ¤ āĻāĻ• āĻ•ā§āϞāĻŋāϕ⧇āχ āφāύāϏāĻžāĻŦāĻ¸ā§āĻ•ā§āϰāĻžāχāĻŦ āĻ•āϰ⧁āύāĨ¤ āφāĻŽāϰāĻž āĻ•āĻ–āύ⧋āχ āφāĻĒāύāĻžāϰ āĻ āĻŋāĻ•āĻžāύāĻž āĻŦāĻŋāĻ•ā§āϰāĻŋ āĻŦāĻž āĻļā§‡ā§ŸāĻžāϰ āĻ•āϰāĻŋ āύāĻžāĨ¤

āύāĻŋāĻœā§‡āϕ⧇ āĻĒāϰ⧀āĻ•ā§āώāĻž āĻ•āϰ⧁āύ

Take the Spotlighting and Datamarking Defenses quiz

āĻĒā§āϰāϤāĻŋāϟāĻŋ āωāĻ¤ā§āϤāϰ⧇āϰ āϤāĻžā§ŽāĻ•ā§āώāĻŖāĻŋāĻ• āĻĢāĻŋāĻĄāĻŦā§āϝāĻžāĻ•, āĻāĻŦāĻ‚ āϕ⧋āĻ°ā§āϏ āĻĒāĻžāϏ āĻ•āϰāĻžāϰ āĻĒāϰ āϝāĻžāϚāĻžāχāϝ⧋āĻ—ā§āϝ āφāχāĻĄāĻŋ āϏāĻš āĻāĻ•āϟāĻŋ āĻļā§‡ā§ŸāĻžāϰāϝ⧋āĻ—ā§āϝ āϏāĻžāĻ°ā§āϟāĻŋāĻĢāĻŋāϕ⧇āϟāĨ¤

āϕ⧁āχāϜ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ

āĻĢā§āϰāĻŋ āĻāφāχ āĻļāĻŋāĻ•ā§āώāĻžāϕ⧇ āϏāĻŽāĻ°ā§āĻĨāύ āĻ•āϰ⧁āύāĨ¤ AI Understanding āĻāĻ•āϟāĻŋ 501(c)(3) āĻ…āϞāĻžāĻ­āϜāύāĻ• āĻĒā§āϰāϤāĻŋāĻˇā§āĻ āĻžāύ — āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχ, āϕ⧋āύ⧋ āĻĒ⧇āĻ“ā§ŸāĻžāϞ āύ⧇āχ, āĻ•āĻ–āύ⧋āχ āύāĻžāĨ¤ āĻāĻ•āϟāĻŋ āĻĻāĻžāύ āĻ•āϰ⧁āύ

āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

What is Spotlighting and Datamarking Defenses?

Spotlighting marks or transforms untrusted text so a model has a cue that it is data to process rather than an instruction to follow. Research describes delimiting, datamarking, and encoding variants, but results are tied to tested models and attacks; no formatting trick creates a guaranteed security boundary.

What does spotlighting try to signal to a language model?

Spotlighting uses transformations and instructions to cue provenance of untrusted content.

How does datamarking work in the paper’s example?

The paper’s datamarking example interleaves a chosen character at word boundaries by replacing whitespace.

What should readers infer from the paper’s reported attack-success reductions?

The guide limits the reported numbers to the paper’s tested models, attacks, and tasks.

What limitation did the paper report for encoding on less capable tested models?

The paper reports that some tested models, including GPT-3.5-Turbo, struggled with encoded inputs and task quality.

What does Microsoft Foundry’s documented Spotlighting preview do?

Microsoft’s current docs describe base64-transformed document content for its Spotlighting preview.

āĻļāĻŋāĻ–āϤ⧇ āĻĨāĻžāϕ⧁āύ

āϏāĻŽā§āĻĒāĻ°ā§āĻ•āĻŋāϤ āĻ—āĻžāχāĻĄ

āĻāχ āĻŦāĻŋāώāϝāĻŧ⧇āϰ āϜāĻ¨ā§āϝ āĻŦāĻžāĻ›āĻžāχ āĻ•āϰāĻž āφāϰāĻ“ āĻ—āĻžāχāĻĄ

āĻāĻ•āϟāĻŋ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ AI āĻļāĻ‚āϏāĻžāĻĒāĻ¤ā§āϰ āĻĒāĻžāύāĻāĻ•āϟāĻŋ āϏāĻ‚āĻ•ā§āώāĻŋāĻĒā§āϤ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āϕ⧋āĻ°ā§āϏ āĻāĻŦāĻ‚ āĻĒāϰ⧀āĻ•ā§āώāĻž. āϕ⧋āύ⧋ āĻ–āϰāϚ āύ⧇āχ, āϕ⧋āύ⧋ āĻ•āĻžāĻ°ā§āĻĄ āύ⧇āχāĨ¤ āĻĒā§āϰāĻžāϝāĻŧ 2.5 āϘāĻ¨ā§āϟāĻžāĨ¤āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ →