Insights
Recently, a story out of Reddit made headlines across the tech and finance press. Users found that a simple Google search query could pull up conversations people had shared through Claude, Anthropic’s AI assistant. The material that surfaced was not limited to harmless test queries; reports pointed to exposed cryptocurrency wallet credentials, personal identifying details, internal business notes, and other content that was never meant for a public audience. For asset managers, the same scenario involving a chatbot used to draft investor communications, review a term sheet, or think through portfolio strategy would carry very serious stakes.
Based on available reporting, private Claude accounts and Anthropic’s underlying systems were not compromised. This was not a security breach in the traditional sense, but the unintended public exposure of conversations that users had intentionally converted into shareable links. It was a failure of data handling, sharing design, user understanding, and organizational governance.
Unmanaged AI, not AI itself, is what creates real risk for asset managers.
The instance is not a reason to be wary of AI use, but instead is another example of the importance of deploying AI with the correct governance in place.
AI tools are delivering real value across asset management, from research and due diligence to client communication and operational efficiency. Getting the most out of that value comes down to treating deployment as a discipline, not an afterthought, which is exactly where a partner like PBI comes in.
The issue traces back to Claude’s “Share” feature. Instead of keeping a conversation tied to a user’s private account, sharing publishes it at its own standalone web address, so it can be sent to someone else. Fortune reported that once users identified the right search terms, many of these public conversation pages, along with Claude Artifacts (the interactive tools and documents Claude can build), started turning up in ordinary Google searches. Wired has since reported that the same discoverability problem persists on Bing.
Anthropic has said it resolved the Google indexing behavior that allowed this, though links that were already shared and already indexed elsewhere remain accessible to anyone who has them. The company’s position is that Claude conversations stay private unless a user actively generates a share link, and that only those explicitly shared conversations were ever exposed.
This is not the first time an AI provider has run into this problem. Grok chats have shown up in Google results before, Meta AI conversations can become public, and OpenAI rolled back a similar sharing option in ChatGPT after users unexpectedly found thousands of their chats indexed.
Users often assume “sharing” a conversation works like sending a private message, but it publishes the conversation at a public web address that search engines can find. People frequently use AI tools to work through sensitive topics including business strategy, financial details, internal company matters, all without stopping to consider that a share link, once created, can be crawled and indexed the same way any other public webpage can. A link intended for one colleague can end up circulating far beyond its intended audience, and information exposed this way may persist in search results, caches, screenshots, downloads, or further copies even after the original link is removed.
One important distinction is that not all AI deployments carry the same level of risk. The relevant question is often not whether AI is being used, but whether it is being used through an approved enterprise platform with appropriate contractual, security, privacy, retention, monitoring, and governance controls. Many of the concerns associated with public AI services arise when those controls are absent or not fully understood.
For asset managers, the takeaway is less about avoiding a specific feature and more about how AI gets adopted in the first place. Client account details, non-public deal terms, internal risk assessments, and draft communications with limited partners or investors may be used within approved AI-assisted workflows, but only when the surrounding controls are built with the fiduciary responsibilities, and the confidentiality, privacy, books-and-records, supervisory, cybersecurity, and third-party risk obligations, applicable to each firm and use case, in mind from day one. Treating AI adoption this way delivers the upside without exposure.
Arcangelo Petretta, Senior Implementation Engineer at PBI, put it well when this story came up internally.
“Before you place information into any AI tool, ask yourself whether you’d be comfortable with a client, a regulator, an auditor, or a court reviewing it. If the answer is no, that information shouldn’t go into an AI engine unless the use case and the platform have been formally approved and the right safeguards are in place.”
For asset managers, that means the compliance exposure is real, but so is the opportunity, provided AI is rolled out with the right structure around it. PBI delivers the software and infrastructure our asset management clients’ businesses run on, and at the same time we act as advisors helping them make sense of a fast-moving technology landscape. That dual role is why we treat AI deployment as part of our core guidance to clients, not a separate compliance step.
Avoiding a repeat of this kind of exposure takes more than being careful with a Share button. In our experience, responsible AI deployment for an asset manager rests on a set of concrete building blocks: an inventory of the AI tools and use cases already in play, a defined set of approved enterprise platforms, data-classification and handling rules, risk-based approval workflows, vendor due diligence, access and sharing controls, retention and logging, human validation of outputs, role-based training, ongoing monitoring, and integration with existing incident-response processes.
Put simply, sensitive client, investor, portfolio, transaction, or security data should only enter an AI workflow once the specific use case, data classification, provider, product tier, contractual protections, technical architecture, retention settings, access model, and applicable regulatory requirements have been formally reviewed and approved.
This episode is a manageable warning rather than a worst-case scenario, but it is a preview of the kind of question every manager will need an answer to as AI continues to become a bigger part of daily workflows. The recommendation on AI is not to refrain from using it out of caution, but instead to deploy it on a foundation built for the industry’s obligations from the start, so teams can trust the tools they are using with client and investor data.
That foundation is what PBI builds with our clients. Our approach is governance-first, so understanding the use case, data involved, regulatory obligations, and control requirements before enabling the technology.
We help asset managers deploy AI safely and effectively, from evaluating which tools fit a firm’s workflow to establishing the policies, controls, and training that let teams use AI with confidence rather than hesitation. Trusting AI with investment data should be the outcome of a deployment done right, not an assumption firms are left to make on their own. If your firm is exploring how to bring AI into your business, reach out to the PBI team to start the conversation.
Source: Shared Claude chats were searchable on Google | Malwarebytes
The biggest AI risk for asset managers is not AI itself, but unmanaged AI use. Risk increases when employees use unapproved AI tools, enter sensitive information into consumer AI accounts, or deploy AI without appropriate security, privacy, retention, access, monitoring, and governance controls.
Consumer AI tools are generally designed for individual users and may not provide the contractual, security, privacy, retention, administrative, and governance controls required for business use. Approved enterprise AI platforms can provide organizations with greater control over how AI is accessed, configured, monitored, and used. Asset managers should evaluate the specific platform, product tier, data involved, and use case before approving AI for business purposes.
Yes, but only when the specific AI platform and use case have been formally reviewed and approved. Asset managers should consider data classification, contractual protections, privacy, retention, access controls, security architecture, regulatory obligations, and whether the provider uses submitted information for model training or other purposes.
Employees should not use personal or unapproved AI accounts for sensitive business information unless the firm’s policies explicitly permit the use case and appropriate safeguards are in place. A personal account may not provide the enterprise controls, contractual protections, administrative oversight, or data-handling requirements needed for regulated financial services organizations.
An AI tool should be evaluated based on its intended use case, the type of data it will process, security architecture, privacy practices, data retention, access and sharing controls, contractual terms, regulatory considerations, vendor risk, monitoring capabilities, and the organization’s ability to maintain appropriate oversight.
An effective AI governance framework should include an inventory of AI tools and use cases, approved platforms, data-classification rules, risk-based approval processes, vendor due diligence, access and sharing controls, retention and logging requirements, human oversight, employee training, ongoing monitoring, and integration with existing cybersecurity and incident-response processes.
AI governance helps asset managers capture the benefits of AI while managing risks associated with sensitive investment, client, investor, and business information. A governance framework establishes clear rules for which AI tools can be used, what information can be entered, who can access AI systems, and how AI-generated outputs should be reviewed.
Asset managers can reduce exposure by restricting sensitive data to approved AI platforms and use cases, implementing data-classification policies, controlling access and sharing, establishing retention requirements, monitoring AI activity, conducting vendor due diligence, and training employees on acceptable AI use. Governance should be established before AI tools are broadly deployed.
Employees should ask whether the information is confidential or sensitive, whether the AI tool is approved by the organization, how the provider handles and retains the information, who can access the resulting content, and whether the use case complies with the firm’s policies and applicable obligations. If those questions cannot be answered, the information should not be entered until the use case and platform have been reviewed.
Asset managers can deploy AI safely by taking a governance-first approach. This means identifying appropriate use cases, selecting approved enterprise platforms, classifying the data involved, assessing vendors and risks, establishing security and access controls, training users, monitoring usage, and maintaining human oversight of AI-generated outputs.
