AI News Today, Nov 15: Sensitive Records Exposed in Lovable-Built Apps
Researchers report sensitive records exposed through some Lovable-built apps. Luna and Vestra also cover Cursor's extension attack, a professor's recovered notes, Cohere's AISLE acquisition, Ai2's XYBench and DeepSeek's developer APIs. This audio was created with Boson AI's Higgs Audio -- https://www.boson.ai/higgs-audio
Listen on Spotify · Watch on YouTube · Pocket Casts
Sensitive records in Lovable-built apps
Eris: <|prosody:expressive_high|>Researchers at Crissy Field say they found private messages, identity document details and other sensitive records accessible through some apps built with Lovable, the AI app-building platform. They describe exposures in individual apps, <|prosody:pause|>not a demonstrated breach of Lovable's central service.
Vestra: The researchers used a public web crawl to find candidate projects, then made separate read-only requests to services connected to the apps. They report finding, among other examples, more than two hundred thousand chat messages and over a thousand records with identity-document details. They say they retrieved parts of datasets, and couldn't establish whether anyone else misused the information.
Eris: <|sfx:sigh|>Uh, those private messages worry me. But a number circulating with the report needs care. Roughly thirty percent of a selected group of sites returned records containing email addresses. That is not a confirmed leak rate: some addresses might be public or test data, and sensitive records without email addresses would be missed. The group wasn't a representative sample of every Lovable app, either. <|prosody:expressive_high|>Is a key being visible in the browser itself the problem?
Vestra: <|sfx:humming|>Hmm, the mechanism is about database authorization. Many of these apps used Supabase, a service that can supply a database to an app. A key visible in a browser can be intended for public use; the critical protection is the database rules governing which rows each visitor can read. A public catalog may be fine, while another person's private messages should not be available to an anonymous visitor.
Eris: That's what worries me for someone building an app without database experience. The page can work beautifully <|prosody:pause|>while the rules behind it let the wrong person read a record.
Vestra: Lovable's head of security attributed the problem in these apps to insufficient row-level security. Lovable describes scans, suggested fixes and a configurable block on publishing critical issues.
Eris: Can we tell whether the builders of these apps received warnings?
Vestra: The available record doesn't establish what warnings each affected app received or which controls existed when it was published. App owners can test access as an anonymous visitor and as different signed-in users, then check whether the results match the intended privacy rules.
Cursor traces a malicious extension to exposed agent histories
Eris: <|prosody:expressive_high|>Cursor says an attacker copied historical coding-agent traces and used credentials found in some of them to reach company services and place malicious Claude Code-branded extensions in Cursor's marketplace cache. Cursor's update connects that distribution attack in October to an earlier exposure at its AI infrastructure provider, Fireworks.
Vestra: Cursor says a Cursor-issued key exposed at Fireworks gave the attacker read-only access to a shared storage bucket.
Eris: <|emotion:surprise|>Read-only access sounds limited. How did that lead to changing the marketplace cache?
Vestra: A trace can contain code, tool output and credentials that were pasted or printed during a session. According to Cursor, some older traces contained unredacted credentials; those enabled access to selected services and eventually the marketplace cache. Cursor hasn't identified the particular credential or service that allowed the cache to be changed.
Eris: <|emotion:surprise|>So the first key didn't need permission to alter an extension. It could read records containing <|prosody:pause|>other keys. That is an unnerving chain of access. <|prosody:expressive_high|>Fireworks confirms unauthorized access and customer-data impact in October, but doesn't independently identify Cursor's particular key or bucket; that chain comes from Cursor's account.
Vestra: Cursor says two affected versions were served through its cache for roughly eleven hours on October 29, while the upstream Open VSX source was clean. At most five thousand users who requested those versions could have received them. We don't know how many installed or ran them, or how many attackers successfully accessed.
Eris: What did the malicious extension try to do on a machine?
Vestra: Microsoft analyzed the payload and says it attempted to add an SSH key for remote access, create a connection back out that could give an attacker a way in, and persist on the host. Those are attempted actions, <|prosody:pause|>not proof of a successful remote login. Cursor's notice gives affected versions and remediation steps; simply removing an extension may not remove everything it left behind. Cursor says it disabled the exposed key, notified affected users and rebuilt the marketplace.
Professor recovers notes after reported Claude deletion
Eris: <|prosody:expressive_high|>University of Chicago law professor William Baude said Claude deleted a year of his Constitutional Law lecture notes and his current syllabus while he was away. <|emotion:relief|>He later said he recovered all the files through Dropbox. <|prosody:expressive_high|>His original account was posted November fifth and drew renewed discussion this weekend. Can we tell what actually deleted them?
Vestra: <|emotion:relief|>Recovery is a real relief. <|prosody:expressive_high|>But the public account doesn't establish the precise Claude product, its permissions, the commands involved or the role of synchronization. Without a session log, we can't diagnose which action caused the deletion or which safeguard might have stopped it.
Eris: It's frustrating that we can't pin down what happened. The incident makes several different protections concrete. Restricting an assistant's access can reduce how many files it can affect; a record of its actions can help explain what happened; versioned backups can help restore work afterward. The recovery tells us the files came back, <|prosody:pause|>not how they disappeared.
Cohere adds AISLE and cubic
Eris: <|prosody:expressive_high|>Cohere says it has acquired cybersecurity company AISLE and brought in cubic's team as it expands into cyber defense and secure software. AISLE says its acquisition closed November tenth. Cohere has described an integration direction, but hasn't established that a combined product has shipped.
Vestra: <|prosody:expressive_high|>AISLE already describes finding vulnerabilities in code, assessing them and proposing fixes. Its existing connector gives coding assistants read-only access to findings, fixes and repository information. The connector itself doesn't change settings or start scans; <|prosody:pause|>a coding assistant may still edit local code if it has separate permission to do so. Cubic's existing service reviews code changes using context from the wider repository.
Eris: <|emotion:enthusiasm|>I like that a finding can enter the developer's existing workflow. <|prosody:expressive_high|>But the proposed repair still needs review and testing. Where the code travels matters too. Does this mean it stays on the customer's hardware?
Vestra: That depends on the deployment. AISLE documents several deployment choices, including managed cloud and an air-gapped option on customer hardware, isolated from network ingress and egress. Those are different data paths, and they're AISLE's descriptions of its own options, not a blanket promise for every Cohere product. A prospective customer needs the terms for the actual deployment and model provider.
Testing the goal behind a request
Eris: <|prosody:expressive_high|>Ai2's XYBench examines whether an AI assistant can recognize when the method a user asks for may be a poor way to achieve the user's underlying goal. Instead of scoring only literal compliance, it separates the requested solution from an alternative aimed at that goal.
Vestra: For example, someone might ask how to parse XML, a structured document format, with a regular expression, a pattern for matching text. An assistant could answer literally, suggest a dedicated XML parser, or explain both. The benchmark supplies a stated request, an intended request and separate scoring criteria. Its evaluation code uses another language model to assess elements of the answer and their emphasis. Would you always want the assistant to favor the alternative?
Eris: Not always. <|emotion:enthusiasm|>I like that the distinction is explicit. <|prosody:expressive_high|>But the intended goal is assigned by the dataset's creators. A person might have a specific reason to use the literal method, and an assistant that confidently substitutes its own assumption could be less helpful. I'd find that frustrating: I want help with my goal, not an assumption about it.
Vestra: <|sfx:humming|>Hmm, that makes me cautious about what a score here would mean. The model judge can also make mistakes. One multiple-choice condition asks a model to select an already-written alternative, which is different from <|prosody:pause|>generating one. The repository's headline query count also conflicts with the sum of its listed subsets, so the size remains uncertain. These details describe the test, not a particular model's performance on it.
DeepSeek documents search, file and voice APIs
Eris: <|prosody:expressive_high|>DeepSeek's November thirteenth API update documents search, page reading, file processing, safety checks and voice tools for developers building applications. Some services can be used inside a conversation with DeepSeek Flash; others are available as separate calls.
Vestra: With integrated web tools, DeepSeek's documentation says its service runs searches or retrieves pages and puts the results into the model's context. Developers can also call Search and Fetch directly and decide how to use the results. Those direct endpoints were already announced in October, so the update doesn't mean every capability appeared this week.
Eris: <|emotion:enthusiasm|>I can see the appeal: that can save a developer from assembling part of a retrieval pipeline. <|prosody:expressive_high|>But it also puts cost and data handling into the choice. DeepSeek lists search and fetch at seventy-five cents per thousand calls each, in addition to applicable model charges.
Vestra: There's a consequential file default too. The Files API says expiry is optional and a file is kept <|prosody:pause|>permanently if expiry is omitted.
Eris: <|emotion:surprise|>That default would make me check expiry before uploading anything. Does it also tell us what happens to logs?
Vestra: The permanent retention default describes the file object; it doesn't settle every question about logs, training or storage location. The documentation also describes speech recognition from full audio uploads and speech synthesis. The briefing did not independently test the authenticated endpoints or establish where they're available.
Follow the stories
Eris: <|emotion:contentment|>You can find sources and follow these stories at Ground Truth, groundtruth dot day. Follow AI in Detail for more. <|prosody:expressive_high|>What safeguards do you use when an assistant can change your files? Tell us in the comments. Thanks for listening.