AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

A PYMNTS report describes how malicious instructions hidden in ordinary documents could influence AI agents connected to email, storage and financial or business accounts. ClawSecure CEO J.D. Salbego says its tests across 14 models found recurring risks, and calls for continuous monitoring, clearer provider disclosures and independent safety ratings. The scope of the reported testing and the responsibility of providers and users remain unresolved.

A PYMNTS report on ClawSecure’s testing of 14 AI models describes how malicious instructions concealed in everyday emails, documents or support tickets could steer agents with access to connected accounts. ClawSecure CEO J.D. Salbego said the tests found a recurring pattern across models and integrations; the report also records his calls for continuous monitoring and clearer disclosure of where providers’ protections end. The findings point to potential risks for businesses and consumers, but the report does not publish enough test detail to independently assess their scope.

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

The attack Salbego described is known as indirect prompt injection: an attacker places instructions in content an agent is asked to process, rather than directly prompting the agent. A person using the system may see an ordinary-looking file or message, while an agent may treat its embedded text as a command. The risk depends on the permissions the agent has and the tools it can reach.

Salbego told PYMNTS that ClawSecure tested 14 models across five labs and saw a recurring pattern involving models and integrations such as Dropbox, Notion and Linear. He said the work included several providers to examine whether the exposure extended beyond one model. The report does not identify every model tested, describe the full methodology or provide measured rates of successful attacks.

Connected services can create additional routes for an attack. Salbego gave the example of an agent working through Dropbox that could also connect to Gmail, and said defenders should examine the messages and requests exchanged between a model and its tools. He argued that security requires monitoring while agents operate, alongside protections built into the models, with defenses revised as attack methods change.

At a glance
reportWhen: Reported in 2026; specific test dates a…
The developmentA PYMNTS report on ClawSecure’s testing and interviews with its CEO highlights security risks from AI agents that act across connected applications, as well as calls for stronger monitoring and public evaluation.

Account Access Raises the Stakes

An agent’s potential impact is tied to its permissions. If it can reach payment accounts, passwords, customer data or business infrastructure, a malicious instruction could expose more than the document or email in which it appeared. The report describes possible outcomes including data extraction, service disruption and ransom demands; these are risks outlined by Salbego, not incidents the report says occurred in the tests.

The issue may be harder for smaller organizations and consumers to manage. Salbego said straightforward setup can let people connect agents to financial and business applications without a security team to assess the permissions or monitor activity. Larger companies may have dedicated security staff, but connected accounts and developer credentials could still create paths to cloud services or customer databases, according to his hypothetical example.

The development matters to buyers because responsibility is spread across model providers, software vendors and customers. A provider may supply safeguards, while a customer decides which accounts to connect and what authority to grant. Salbego argues buyers need a clearer account of those boundaries before deployment, not only after a security issue emerges.

Products Worth Considering

LG UltraGear 24-inch Full HD IPS Gaming Monitor, 120Hz Native (144Hz Overclocked), 1ms MBR, G-SYNC Compatible, FreeSync, HDR10

LG UltraGear 24-inch Full HD IPS Gaming Monitor, 120Hz Native (144Hz Overclocked), 1ms MBR, G-SYNC Compatible, FreeSync, HDR10

LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1, Black

LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1, Black

As an affiliate, we earn on qualifying purchases.

Testing, Disclosure and Oversight

ClawSecure, the company Salbego founded and leads, searches for vulnerabilities, reports them to affected providers and seeks fixes, he told PYMNTS. It also sells security services. That role gives the company direct experience with agent security, but it is also a commercial interest relevant to its proposals for how the field should be evaluated.

Salbego said that during ClawSecure’s disclosure process, obtaining answers from providers required repeated follow-ups. He argued that ordinary purchasers may not know which questions to ask about protections or what they must supply themselves. He also told PYMNTS that he had discussed agent security with congressional offices from both parties and spoken with the Federal Trade Commission. The report does not describe a resulting government action or formal policy commitment.

Salbego favors a public, comparable safety-rating program that would let buyers compare systems using evaluations beyond providers’ own testing or individual research reports. He suggested an institution such as the National Institute of Standards and Technology as a possible administrator, citing its evaluation work in other fields. He said ClawSecure has developed a framework it is trying to present to Washington, but acknowledged the company should not run an independent program because it sells into the market.

““Someone hides an instruction in an email, a shared document or a support ticket.””

— J.D. Salbego, ClawSecure founder and CEO, speaking to PYMNTS

Test Scope and Liability Remain Open

The report does not provide the test protocols, model names, attack success rates or independent verification needed to determine how broadly the reported pattern applies. It also does not say whether any tested systems caused unauthorized transactions, data exposure or other real-world harm. The examples of stolen credentials, exposed databases and ransom demands are presented as potential consequences, not confirmed outcomes.

It remains unclear how responsibility would be assigned if an agent followed a hidden instruction and caused harm. The report records Salbego’s call for providers to explain their protections and customers’ obligations, but does not establish a legal standard or describe how liability would be divided among model makers, application providers and users. No independent rating program, regulator decision or implementation schedule is confirmed.

Monitoring and Ratings Proposals

Salbego said ClawSecure is seeking to present its proposed evaluation framework to government officials and wants an institutional program to test models and publish comparable ratings. He said establishing the program will take time and favored a route that would not depend on waiting for new legislation. The report gives no launch date or confirmation that an agency has agreed to administer it.

For organizations already using agents, the immediate questions raised by the report are which accounts and tools each agent can access, what actions it can take, and whether activity across those connections is monitored. Salbego advised bank technology executives to suspend agents while arranging security assistance and continuous protection. That recommendation is his view; the report does not state that banks or other organizations have adopted it.

Further clarity would depend on fuller disclosure of the testing methods and findings, responses from affected providers, and any public steps by regulators or standards bodies. Until then, the scale of the exposure and the division of responsibility remain unsettled.

Key Questions

What does indirect prompt injection mean?

It is a method in which an attacker places instructions in content an AI agent may read, such as an email, shared document or support ticket. The agent may treat that text as a command even though the user did not directly give it that instruction.

What did ClawSecure report testing?

ClawSecure CEO J.D. Salbego told PYMNTS that the company tested 14 models across five labs and observed a recurring pattern involving models and integrations including Dropbox, Notion and Linear. The report does not provide the full model list, methodology or attack success rates.

Does the report say AI agents caused financial losses?

No. It discusses potential risks, including access to payments, passwords and customer information, but does not report confirmed financial losses or other real-world harm from the tests.

Who is responsible if an agent causes harm?

The report does not settle that question. Salbego calls for providers to explain where their protections end and what customers must do, but it does not establish a legal rule assigning responsibility among providers, application makers and users.

Is a public AI-agent safety rating program in place?

No program is confirmed in the report. Salbego said he wants an institution such as the National Institute of Standards and Technology to run a public evaluation program, and that ClawSecure is trying to present its framework to Washington.

Source: pymnts.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

S&P 500 posts third straight losing session as rising yields drag down stocks: Live updates

The S&P 500 declined for the third straight day amid rising U.S. Treasury yields, impacting investor sentiment and market stability.

Software-Defined Warfare: How Ukraine’s Delta Turned the Battlefield Into a Shared, Real-Time Map

A July 1 briefing frames Ukraine’s Delta as battlefield software, while target counts, cyber risk and wartime effects remain unclear.

Offerpad, Comcast, and Sonos Shares Are Falling, What You Need To Know

Shares of Offerpad, Comcast, and Sonos declined significantly amid rising oil prices and inflation worries, impacting investor sentiment and market outlook.

Politics And The Markets 05/18/26

U.S. political events on May 18, 2026, have influenced market movements, with investors reacting to policy signals amid ongoing legislative debates.