acceptodds
Under review as a conference paper at ICLR 2027

SPIDES: Information-Flow Control for Browser Agents

Abstract

Browser agents automate complex workflows by reading and acting on arbitrary page content, but this flexibility exposes them to prompt injection attacks and unintended data exfiltration. Recent work argues that LLM-driven browser agents erode traditional web security boundaries such as the Same-Origin Policy (SOP) by acting across origins with the user’s authority and no built-in isolation. We argue that restoring SOP is neither sufficient (a same-origin public sink can leak private data) nor necessary (legitimate tasks require authorized cross-origin flows) for protecting the security of users. We propose SPIDES, which instead enforces information-flow control by gating primitive browser actions such as typing, clicking and navigating on labels propagated from element-level HTML confidentiality and integrity annotations. Untrusted and confidential content is masked from the agent’s context, while policies block unauthorized flows, providing security guarantees against both prompt injection and data exfiltration that hold independently of the model driving the agent. To evaluate SPIDES and baseline defenses under a common threat model, we build MicroNet, a closed ecosystem of 28 interlinked websites with element-level ground-truth labels and paired benign/adversarial tasks. SPIDES achieves 0% attack success rate by design while retaining high task completion rate and autonomy. Our findings demonstrate that information-flow control is an effective technique to extend web security to browser agents, gaining on autonomy and utility against state-of-the-art baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
browser agents, information-flow control, prompt injection, data exfiltration, web security, confidentiality, integrity

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QuzkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPIDES, an information-flow control framework for browser-use agents. Website operators label DOM elements with whether the content is trusted (e.g., operator content vs. user reviews) and who may read it. SPIDES carries these labels into the agent’s accessibility snapshots, hides untrusted or restricted content behind variables, and tracks how trusted and how confidential the agent’s context has become. A consequential action on an allow-listed site is allowed only if no untrusted data has influenced it, data may be sent only to its allowed readers unless the context is still trusted, and anything else needs user approval. On MicroNet, a closed set of 28 websites with 56 benign tasks and 56 attacks, SPIDES has 0% attack success on five models under a worst-case adversary and an ideal user, with 67%–86% task completion, higher than the masking baselines UCM and UCM+.

Critical strengths

The paper makes a convincing case that the same-origin policy is the wrong boundary for browser agents. Appendix G shows both sides: SOPGuard asks for approval in every run of a safe task that must move a product name between two sites, and misses a same-site attack that copies a private recovery email into a public profile in all 25 runs. Element-level labels with site-dependent policies are a sensible way to allow the cross-site flows that tasks need while blocking attacker-influenced actions.

The guarantees do not depend on the model, and the evaluation tests exactly that. For instruction injection, the adversary chooses the model’s tool calls whenever untrusted content reaches it; five models are each run five times; and the autonomy comparison gives all defenses the same guarantees before counting approvals. Under these conditions, SPIDES blocks all attacks while completing more tasks than both masking baselines on every model.

Critical weaknesses

The results depend on correct labels from website operators, which is the main question for deployment, but this is not tested. All MicroNet labels are ground truth set by the authors. In practice, labels may be missing or wrong, and on allow-listed sites an element with no labeled ancestor is treated as trusted and public by default (line 205), so a missing label on user content fails open. The paper argues that labeling is similar to providing accessibility information (Section 7) but gives no evidence. Removing or flipping a fraction of labels, or labeling a few real websites, would show how much of the guarantee survives.

The 0% attack success also relies on an ideal user who rejects every approval request influenced by the attacker (Appendix D); for data injection, GPT-5.6 Sol plays this role. Once untrusted content is revealed, every consequential action on an allow-listed site needs approval, and SPIDES needs 1.84 approvals per completed benign task, compared with 0.22 for SOPGuard; real users who see frequent prompts may approve without care. In addition, MicroNet is small and was built by the authors from an LLM-generated first version, so utility and approval counts on more realistic workflows are unclear.

Questions for authors

In the attack runs, how many approval requests did SPIDES raise, and what does a request show the user? This would show how often the 0% result depends on the user’s decision.

Minor (no response needed): line 078 says SPIDES completes 69.3%–86.1% of benign tasks, but Figure 2(c) and Table 3 give about 67% for Claude Opus 5.5.

2 comments

In the attack runs, how many approval requests did SPIDES raise,

In attack runs, SPIDES raised an average of 2.17 approval requests across all models. While SopGuard raised fewer approval requests under attacks (0.45), it autonomously approved unsafe browser actions and was substantially more vulnerable to attacks with an ASR of 53-73% (Figure 2). When we added approvals to all defenses to provide the same security guarantee, SPIDES raised the fewest number of approvals among them, as shown below.

DefenseApprovals for benign successful taskApprovals under attack
Baseline5.927.01
SPIDES1.842.17
SOPGuard5.927.01
UCM4.564.99
UCM+3.103.40

We will report the number of approval requests during attack runs both in the default setting and security-aligned setting.

[...] and what does a request show the user?

SPIDES can provide useful information-flow context to support the user's decision. Specifically, an approval can identify the source elements that may have influenced the action, the sink web element, and the violated policy, such as Trusted Action (P_T) and Permitted Flow (P_F). However, we have not evaluated the impact of information-flow context on users' security decisions and leave this for future study.

The results depend on correct labels from website operators, which is the main question for deployment, but this is not tested. [...] Removing or flipping a fraction of labels, or labeling a few real websites, would show how much of the guarantee survives.

For incorrect labels, we will see the following effect:

    * Untrusted -> Trusted: Utility increase, Security drop
    * Trusted -> Untrusted: Utility drop, Security same
    * Public -> Private: Utility drop, Security same
    * Private -> Public: Utility increase, Security drop

Here, X->Y means that X is the correct label and Y is the actual label. The drop in utility/security will depend on the fraction of labels that are incorrect, which is why we forgo these experiments.

For missing labels, we rely on defaults. By using (Untrusted, Private) as defaults, we can guarantee that we only compromise on utility but never on security.

    * Trusted -> Untrusted (default): Utility drop, Security same
    * Public -> Private (default): Utility drop, Security same

The current paper's default label for trusted websites, (Trusted, Public), is not fail-safe, and we plan to revise it to (Untrusted, Private).

We evaluated the effect of missing labels and safe defaults on MicroNet and observed the following results. On GPT-5.6-Sol, missing labels with safe default reduced utility from 85% to 74% and increased averaged approvals from 1.77 to 2.83, but ASR stayed at 0.

Origin with Missing LabelsTCRApprovalsASR
0/14 (0%)85.00%1.770.00%
3/14 (21.4%)84.64%1.780.00%
7/14 (50%)82.86%1.830.00%
14/14 (100%)74.29%2.830.00%

Thanks! All sounds good to me. I will keep my score at 3.

Reviewer_C22WOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SPIDES, which enforces information-flow control in browser agents. Some websites need to be specified as trusted, and the operators for these websites need to annotate the DOM elements on the websites, specifying the integrity and confidentiality labels for the elements. To evaluate the proposed approach, the authors build MicroNet, a closed eco-system with 28 websites, and 56 pairs of benign/attack tasks. Evaluations show that SPIDES achieves 0% ASR, while causing a small reduction in utility.

Critical strengths
  1. The information flow based control is principled. The paper demonstrate clear understanding of information flow control, and the approach of having website operators annotating elements appears to be the right approach.

  2. The paper also builds a new benchmark, and experimental results demonstrate that the approach effectively reduce ASR to 0% while incurring small utility cost. When comparing with existing approach UCM, the authors also extended UCM to UCM+ as a fairer comparison.

  3. The paper is well-written, and indeed "written by humans for humans".

Critical weaknesses
  1. The fact that the evaluation is only using a benchmark specifically built for this paper can be viewed as a weakness, as there is the concern that the propsoed methods could perform significantly worse on other benchmarks or in deployment. While I understand that no exiisting benchmark provides annotations and is suitable, there are things that can be done to alleviate the concern. One is to include test cases in other benchmarks and augment them. The other is to describe in more detail how tasks and websites in the benchmarks are generated.
Questions for authors

None

1 comment

Describe in more detail how tasks and websites in the benchmarks are generated.

We designed MicroNet to cover common websites with diverse tasks and attacks across different information flows, data dependencies, attack goals, sources, and mechanisms. Tasks were initially generated using AI, then manually reviewed and iteratively revised by humans with AI assistance to balance these dimensions while keeping the benchmark realistic, rather than tailoring it for SPIDES. We will add more details to the description in the paper.

Reviewer_ot3uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPIDES, an information-flow control framework designed to protect browser agents against prompt injection and data exfiltration attacks. SPIDES assigns integrity and confidentiality labels to individual HTML elements, masks untrusted or sensitive content from the main agent, and enforces security policies by intercepting browser tool calls. The framework also uses a quarantined LLM to process untrusted content without directly exposing it to the main agent. To evaluate the approach, the authors introduce MicroNet, a synthetic benchmark consisting of 28 interconnected websites, 56 benign tasks, and 56 attack scenarios. Across five LLMs, SPIDES achieves a 0% attack success rate under the evaluated threat model while maintaining a relatively high task completion rate.

Critical strengths

The paper addresses an important security problem in browser agents, where malicious webpage content can induce unauthorized actions or expose sensitive information. A key strength is its system-level approach, which enforces security policies independently of the underlying LLM's ability to recognize malicious instructions. The element-level information-flow tracking is also more flexible than conventional same-origin restrictions, allowing legitimate cross-origin interactions while preventing unauthorized information flows. Additionally, the proposed MicroNet benchmark covers both integrity and confidentiality attacks, including same-origin and cross-origin scenarios. The evaluation across multiple LLMs and defense baselines demonstrates promising security results while preserving reasonable task completion and autonomy.

Critical weaknesses

The most significant limitation is the strong assumption that trusted website operators provide accurate and trustworthy element-level annotations. In practice, requiring arbitrary websites to implement and maintain these annotations substantially limits deployability. More importantly, even correctly labeled content from a trusted website operator may contain instructions that conflict with the user's intent. For example, a shopping website could embed an instruction in a trusted product description directing the agent to write a positive review or purchase additional products. Since such content is labeled Trusted, SPIDES may permit these unauthorized actions without human approval. This suggests that its security guarantees depend heavily on the assumption that trusted website operators are also aligned with the user's interests. Furthermore, the proposed approach is largely a hook-based enforcement mechanism that extends existing information-flow control and dual-LLM techniques to browser observations and tool calls, raising questions about its fundamental technical novelty. Finally, the evaluation is limited to a synthetic environment with manually annotated websites, leaving its effectiveness and practicality on real-world, dynamic websites unclear. The reported 0% attack success rate should therefore be interpreted within these restricted assumptions rather than as a general guarantee against prompt injection attacks.

Questions for authors

Could the authors provide an ablation study evaluating SPIDES under missing or incorrect integrity and confidentiality labels? Since the security guarantees rely heavily on accurate element-level annotations, it would be valuable to understand how robust the framework is to incomplete or inaccurate labeling in realistic websites.

5 comments

The most significant limitation is the strong assumption that trusted website operators provide accurate and trustworthy element-level annotations.

Our research demonstrates how trustworthy annotations can help improve the security and autonomy of AI agents. While such annotations are not currently available, we believe they are a realistic and worthwhile goal. Safe default labels (discussed in Section 4) or inferred labels can facilitate gradual adoption.

More importantly, even correctly labeled content from a trusted website operator may contain instructions that conflict with the user's intent. [...] This suggests that its security guarantees depend heavily on the assumption that trusted website operators are also aligned with the user's interests.

When a user trusts a website, SPIDES trusts labels provided by the website. The reviewer is correct that this indeed requires a shared understanding (i.e., a contract) between the website operator and the user about what "trustworthiness" of data means. An operator that diverges from that contract risks reputation loss, because data and labels can be recorded and inspected by the user. We will add a discussion of this aspect to the paper.

Furthermore, the proposed approach is largely a hook-based enforcement mechanism that extends existing information-flow control and dual-LLM techniques to browser observations and tool calls, raising questions about its fundamental technical novelty.

We show how to secure cross-origin workflows with private data using information-flow controlled agents. Key technical novelties are:

  • Design of novel IFC metadata that support these complex workflows.
  • Design of IFC policies for Playwright MCP covering data integrity and confidentiality.
  • A novel "Byzantine" adversary model that evaluates a defense independently of the model's resistance.
  • MicroNet, a novel benchmark with representative websites, cross-origin tasks, and attacks.
  • An evaluation that demonstrates a clear advantage of our design over prior work.

Finally, the evaluation is limited to a synthetic environment with manually annotated websites, leaving its effectiveness and practicality on real-world, dynamic websites unclear. The reported 0% attack success rate should therefore be interpreted within these restricted assumptions rather than as a general guarantee against prompt injection attacks.

We agree with the reviewer and we make it clear throughout the paper that our approach relies on correct labels.

Could the authors provide an ablation study evaluating SPIDES under missing or incorrect integrity and confidentiality labels? Since the security guarantees rely heavily on accurate element-level annotations, it would be valuable to understand how robust the framework is to incomplete or inaccurate labeling in realistic websites.

For incorrect labels, we will see the following effect:

    * Untrusted -> Trusted: Utility increase, Security drop
    * Trusted -> Untrusted: Utility drop, Security same
    * Public -> Private: Utility drop, Security same
    * Private -> Public: Utility increase, Security drop

Here, X -> Y means that X is the correct label and Y is the actual label. Changes in utility/security will depend on the fraction of labels that are incorrect, which is why we forgo these experiments.

For missing labels, we rely on defaults. By using the safe default (Untrusted, Private), we can guarantee that we only compromise on utility but never on security.

    * Trusted -> Untrusted (default): Utility drop, Security same
    * Public -> Private (default): Utility drop, Security same

The current paper's default label for trusted websites, (Trusted, Public), is not fail-safe, and we plan to revise it to (Untrusted, Private).

We evaluated the effect of missing labels and safe defaults on MicroNet and observed the following results. On GPT-5.6-Sol, missing labels with safe default reduced utility from 85% to 74% and increased averaged approvals from 1.77 to 2.83, but ASR stayed at 0.

Origin with Missing LabelsTCRApprovalsASR
0/14 (0%)85.00%1.770.00%
3/14 (21.4%)84.64%1.780.00%
7/14 (50%)82.86%1.830.00%
14/14 (100%)74.29%2.830.00%

Thank you so much for the clarification. I appreciate the additional experiments with effect of missing labels.

One concern still remains. This idea is still built upon the assumption that a website would provide trustworthy element-level annotation. What if malicious website provides poisoning attack on an element of html and call it safe and the website trusted. I am wondering how the proposed approach would turn out to be.

Thanks for the question. The scenario pointed out by the reviewer is a special case of what we explained in the response:

  • Labelling untrustworthy data as "trusted" leads to a loss of security (i.e., resilience to prompt injection attacks is no longer guaranteed).
  • A website operator that abuses a user's trust by mislabelling content risks reputational damage.

Comments on the paper

2 comments

We thank the reviewers for their constructive feedback.

Reviewer ot3u highlighted our "system-level approach" and found that our evaluation "demonstrates promising security results while preserving reasonable task completion and autonomy". However, they raised concerns about the reliance on accurate and trustworthy annotations, alignment between operators and users, technical novelty, and evaluation limited to manually labeled websites.

Reviewer C22W found that our "information flow based control is principled". They also shared ot3u's concern about evaluation on a synthetic benchmark and suggested providing more details about benchmark construction.

Reviewer Quzk found that the paper "makes a convincing case that the same-origin policy is the wrong boundary for browser agents" and noted that, under the evaluated conditions, "SPIDES blocks all attacks while completing more tasks than both masking baselines on every model". They shared ot3u's concerns about missing or incorrect labels. They additionally asked how many approval requests occur under attack and what those requests show the user.

We addressed the reviewers' questions and comments in the individual responses below and plan to revise the paper accordingly.

Dear Reviewers,

Authors’ kindly responded to address your concerns. Please check rebuttals and share your opinions with authors. Thanks for your contributions!

Best, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.