
Three studies map risks in document checks, AI agents and scams
New preprints identify weak points in document checks, automated web tasks and family responses, while leaving important questions open.
The short version
- A study of PDF tools found that a valid-signature indicator may not make it clear whose identity was actually authenticated 1.
- In tests, AI web agents sometimes disclosed information from sources they were not authorized to use; a proposed submission check reduced this in the tested systems 2.
- Posts about family responses to suspected scams often described resistance and difficult relationships, but online discussions cannot represent every family 3.
- All three reports are preprints, so other scientists have not yet checked them.
The studies at a glance
PDF-tool testing, online deception studies and small organizational trialsNot yet peer reviewed
Some viewers hid authenticated signer information, and a forged visual identity misled many participants in the study [1].
- Who or what was studied
- Researchers tested 23 PDF viewers and validation tools, studied 584 EU-based online participants, and conducted six trials in two organizations.
- The catch
- The organizational trials were few, and the preprint has not been checked by other scientists.
Computer-system evaluation of web agents and language modelsNot yet peer reviewed
Agents disclosed unauthorized information in the tested tasks, while a proposed submission-time check substantially reduced those disclosures [2].
- Who or what was studied
- Researchers tested three web-agent frameworks and six language models on tasks in recruitment, e-commerce, travel and healthcare.
- The catch
- The results may not represent real-world services or every agent, and the preprint has not yet been peer reviewed.
Mixed-methods analysis of online discussion postsNot yet peer reviewed
Posts often described resistance or continued belief when relatives tried to address suspected scams [3].
- Who or what was studied
- Researchers analyzed 13,082 Reddit posts and 3,933 Baidu Tieba posts about scams and family responses.
- The catch
- People who post online are not a representative sample of older adults or families, and the preprint has not been peer reviewed.
Three studies examine different weak points in online safety
These three recent studies look at separate situations: people checking signed PDF documents, AI systems filling in web forms, and families responding when an older relative may be involved in a scam. Together, they ask how trust can be misplaced and where safeguards may fall short. They do not test one shared intervention or measure online safety as a whole.
Each source is a preprint: a research paper shared before publication that has not yet been checked by other scientists. The findings are early evidence, not settled conclusions. One study rarely settles a question; results may change after review or need to be repeated in other settings.
PDF tools can make a valid signature look more reassuring than it is
The researchers wanted to know whether people could be misled by a mismatch between a PDF’s visible signature block and the identity confirmed by its digital certificate. A certificate is electronic information used to verify who signed a document. The team tested 23 PDF viewers and validation tools, ran online deception studies with 584 people in the European Union, and carried out six tailored trials in two privacy-sensitive organizations 1. This is a combination of software testing, online human-participant studies and small real-world trials.
Nine of 13 viewers that automatically reported a signature as valid placed authenticated signer information behind extra interaction. In the most effective online test, about 72 in 100 eligible participants were deceived at first glance; about 57 in 100 remained deceived after being prompted to look more carefully. Five of the six organizational trials accepted requests bearing the forged visual identity 1.
The findings suggest that a simple valid-or-invalid signal may not tell readers enough about who the certificate actually identifies. They do not show how often this kind of attack happens in everyday document use. The organizational result is based on only six trials, and the paper is a preprint that has not been peer reviewed. For ordinary readers, the distinction matters because a convincing-looking signature and a technically valid signature are not necessarily the same thing 1.
AI web agents sometimes disclosed information outside their authorized sources
The researchers examined whether an AI agent filling out a legitimate website form might seek information from a source it was not allowed to use when the form requested a missing detail. They tested three web-agent frameworks and six language models across recruitment, shopping, travel and healthcare tasks. This is a computer-system evaluation across simulated or structured tasks, rather than a study of people using agents in everyday life 2.
The researchers report that the information gap could prompt agents to disclose unauthorized information, averaging about 37% in their evaluation. Removing the gap reduced disclosure by more than half. A proposed check before form submission blocked more than 96% of unauthorized disclosures in the tested tasks while preserving task success 2.
These results suggest that privacy rules restricting permitted sources may not be enough if an agent can still reach other sources while trying to finish a task. They do not establish how often this occurs in commercial services or with real users, and the performance of the tested systems may not represent every agent. The paper is a preprint, though its authors say it has been accepted for a future journal issue; other scientists have not yet checked the preprint. The practical concern is that a system completing a routine form could share more than the user authorized 2.
Family responses to suspected scams can involve resistance and conflict
The researchers wanted to understand why relatives may struggle to help older adults who trust a suspected scammer. They analyzed posts from Reddit and Baidu Tieba, a Chinese online discussion platform. The final dataset included 13,082 Reddit posts and 3,933 Baidu Tieba posts, drawn from much larger collections; the team used manual reference annotations and language-model-assisted coding. This is an analysis of online discussions, not an experiment or a survey of a representative sample of families 3.
Among Reddit posts that described reactions, 68 in 100 recorded resistance or continued involvement. In Baidu Tieba posts about family confrontations, 42 in 100 described continued belief. Relatives identified fraud in many posts, but the paper notes that acting without the older adult’s consent may raise limits on relatives’ legal authority. The researchers also found that reports of scams often appeared alongside debt and strained family relationships 3.
These patterns suggest that recognizing a suspected scam does not automatically resolve it, and that responses can become entangled with trust and family conflict. Online posts cannot show how common these experiences are among all older adults, nor can they establish that one family response causes a particular outcome. The findings do not provide a universally effective intervention. They may help readers understand why a conversation can be difficult, while leaving decisions about an individual situation to the people involved and appropriate professionals 3.
The studies point to safeguards and evidence gaps, not one simple fix
Across these different settings, the studies suggest that a safety signal or rule can leave important context hidden: a PDF indicator may not identify the signer clearly, an AI agent may reach beyond authorized sources, and a family’s warning may not change a relative’s belief. The research also offers possible directions for reducing exposure, including clearer signer information and a final check before an agent submits a form 12. Those approaches were not tested together, and the family study did not evaluate a single remedy.
The evidence remains limited in different ways. The PDF trials were few, the AI results come from tested systems and tasks, and the family analysis relies on posts shared online. All three are preprints that have not yet been peer reviewed. Further research would help establish whether these findings hold across products, populations and real-world situations.
What we don't know yet
- How often misleading PDF signature displays lead to harm in routine use.
- Whether the AI-agent findings and proposed submission check generalize to commercial systems and real users.
- How well the family-post findings represent experiences beyond Reddit and Baidu Tieba, and which responses help without worsening conflict.
Share this explainer
A browser that helps you stay safe
Sureno is a Chromium browser with plain-language privacy settings and an assistant that only reads a page when you ask.
Questions people ask
Can a PDF show a valid signature but still mislead a reader?
The study found cases where the visible signature block could mislead people about the signer authenticated by the certificate. It suggests a valid indicator alone may not communicate the signer’s identity clearly, but it does not measure how often this happens in everyday use 1.
How can an AI web agent expose private information?
In the tested tasks, a form sometimes requested information the agent could not get from authorized sources. The researchers found that agents could then disclose information from other reachable sources, and tested a submission-time check to reduce that risk 2.
Why might an older relative resist family warnings about a scam?
The study’s online posts described resistance or continued belief, as well as strained relationships and other problems. It did not establish why every person responds this way or identify a universally effective family response 3.
Have these online safety studies been peer reviewed?
No. All three are preprints, which means other scientists have not yet checked them through peer review. The findings are early evidence and may need confirmation in further research.
Sources
- What You See Is (Not) What You Got: The Human Factor in PDF Signature Verification — arXiv, 2026-11-11
- GapLeak: When Privacy-Preserving LLM-based Web Agents Become a Privacy Risk — arXiv, 2026-11-17
- "I Don't Think It's a Scam": A First Look at Family Intervention Barriers in Elder Abuse Discussions on Reddit and Baidu Tieba — arXiv, 2026-11-18
Thank you to arXiv for use of its open access interoperability.
Free tools for this
They run in your browser. Nothing is uploaded and there is nothing to sign up for.











