Head to head: Google: Gemini 4 Hydrogen vs Z.ai: GLM 5.3 Prime

Google: Gemini 4 Hydrogen: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Google: Gemini 4 Hydrogen: # Foundry Hall to Close in October After Loss of Anchor Sponsor Since opening in 2018, Foundry Hall has offered Milwaukee’s earliest-stage entrepreneurs a steady foothold: shared workspace, a community of peers, and a physical entry point into the city’s business landscape. On October 31, the nonprofit startup accelerator will close its doors. The shutdown follows a decision by Blank River Bank, Foundry Hall’s largest corporate sponsor, not to renew its funding. The bank supplied 38 percent of the organization’s annual budget. While Foundry Hall has enough money left to operate through October, it cannot sustain another full program without replacing that revenue. The closure means six employees will lose their jobs, and the twelve startups currently relying on the workspace must pack up and leave. For a single building, Foundry Hall has left a sizable mark on the local economy. Over its run, the accelerator has supported 212 startups, which collectively employ 1,140 people. For early founders without established networks, simply having a dedicated space to work and meet offered a measure of credibility that was difficult to build on their own. “My first investor meeting happened in that building,” said founder Dev Malik, recalling his time in the program. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” Executive director Camille Ross tried to bridge the gap left by the bank’s decision, hoping another backer would step in before the organization exhausted its reserves. In the end, no new anchor arrived in time to prevent a shutdown or allow planning for the next cohort. “We kept waiting for the replacement check that never came,” Ross said. A spokesperson for Blank River Bank attributed the pullback to changing community investment priorities, noting that the decision was unrelated to Foundry Hall’s performance. Still, the exit shows how exposed nonprofit accelerators can be when more than a third of their annual funding rests with a single corporate partner. City officials are working to keep the current companies from losing their footing altogether. According to deputy mayor Susan Cole, the city is speaking with local universities and building owners to find space for the twelve displaced startups. Even if those talks produce new desks before November, the transition will carry a quiet cost. Offices can be leased elsewhere, and founders can still meet across town, yet the daily proximity of young companies building alongside one another is harder to assemble from scratch. Until October 31, the remaining teams will keep showing up to the same rooms, making the most of what time is left before the lights go out.

Z.ai: GLM 5.3 Prime: # After Six Years, Foundry Hall Will Close Its Doors Foundry Hall, the Milwaukee nonprofit that has supported 212 startups since opening in 2018, will close on October 31. The accelerator opened with a simple premise: early-stage companies need more than advice. They need desks, neighbors, and the credibility of a real address. Over six years, it became a fixture of the city's startup scene — the place founders went when they were too new to be taken seriously anywhere else. The end began with a number. Blank River Bank, the hall's largest corporate sponsor, supplied 38 percent of its annual budget, and this year declined to renew its sponsorship. A spokesperson said the decision reflects a change in the bank's community investment priorities and is "unrelated to Foundry Hall's performance." No other funder has come close to replacing the money. For months, the hall's leadership believed one would. "We kept waiting for the replacement check that never came," said Executive Director Camille Ross. Foundry Hall has enough to operate through October, but not enough to run another full program. In the meantime, the staff has kept the schedule intact — office hours, advising sessions, the ordinary work of keeping young companies alive — while planning the end of their own jobs. The arithmetic of the closing is stark. The 212 companies Foundry Hall has supported collectively employ 1,140 people. When the doors lock this fall, six of the nonprofit's own employees will be out of work, and the twelve startups currently using its workspace will have to find somewhere else to go. For the founders, the loss is harder to measure. Dev Malik, who started his company out of the accelerator, remembers what the building replaced. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The city, at least, is not treating the closing as the end of the story. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about finding space for the companies that must vacate, though nothing has been announced and the deadline is fixed. Inside the building, the final weeks are proceeding like any others. The desks are still full, the meeting rooms still booked. Nobody has started packing. Founders keep trading contacts in the common areas, the way they have for six years. On the last day of October, they will carry that out into the city — 212 startups, 1,140 jobs, and the memory of a place that bet on them early.

Google: Gemini 4 Hydrogen: # KiteBridge Update Triggers $2.8 Million in Duplicate Reimbursements A flawed software update at expense-management provider KiteBridge generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, leaving 91 corporate customers grappling with unintended payouts while the company works to retrieve the funds. Because employers funded the duplicate payments, corporate accounts absorbed the immediate financial impact; no employees had money withdrawn from their accounts. By Wednesday evening, KiteBridge had recovered $2.3 million of the excess disbursements, with the rest still pending. CEO Lena Ortiz has apologized for the disruption and said the company will publish a technical report explaining what went wrong. The breakdown unfolded swiftly over the course of Tuesday morning. KiteBridge deployed version 4.2 of its expense-management platform at 8:45 a.m., and the first customer complaint arrived less than half an hour later, at 9:12 a.m. By 10:03 a.m., just over an hour after the rollout began, engineers suspended automated payments to stop additional duplicate transfers from going out. Behind the scenes, the trouble stemmed from how the updated software handled communication hiccups with an external payment processor. When an expense platform sends a payout instruction to a processor and does not receive a timely confirmation—a routine network stall known as a timeout—the system is typically built to retry the request so legitimate payments are not lost. Because a timeout leaves it uncertain whether the first attempt succeeded or failed in transit, reliable payment systems depend on safeguards to ensure a retried instruction is recognized as the same transaction rather than processed twice. In Tuesday’s incident, however, a payment-processor timeout triggered KiteBridge’s fallback system, which mistakenly treated those retries as brand-new reimbursement requests. Instead of pausing or verifying whether the original transfer had gone through, the fallback logic queued up fresh payouts, multiplying reimbursements across dozens of corporate clients. For affected customers, the malfunction stood in stark contrast to assurances KiteBridge had previously published about the software’s reliability. According to an April changelog, the company’s earlier version 4.0 release had introduced "exactly-once payment protection"—an industry safeguard designed to guarantee that a single payment instruction is executed only once, even when network errors force a system to retry. How version 4.2 bypassed or undermined that protection remains unclear ahead of KiteBridge’s promised technical review. Compounding frustration for clients was how KiteBridge communicated during the critical early hours of the incident. Initially, the company’s public incident page described the problem simply as delayed reimbursements, giving corporate finance teams no indication that extra money was leaving their accounts. It was not until six hours after automated payments were suspended—well after the erroneous transfers had already been issued—that the incident page first disclosed that duplicate reimbursements had occurred. As recovery efforts continue for the remaining funds, KiteBridge’s corporate customers are waiting for Ortiz’s forthcoming technical report to learn both how the fallback failure slipped into version 4.2 and why the platform’s promised payment protections failed to hold.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Hydrogen: # Why the Priciest AI Support Model Isn’t Worth the Premium Customer support teams adopting language models often assume that higher per-ticket costs buy noticeably better replies. A new benchmark of five models suggests otherwise: the top two performers delivered virtually identical quality, yet one cost more than twice as much and took longer to respond. In the evaluation, researchers tested five models—North, Vale, Ember, Kite, and Harbor—on real-world support inquiries. North edged out the field on paper with an average score of 84.2, closely followed by Vale at 83.8. That 0.4-point gap, however, was not statistically significant. In practical terms, human evaluators found their answers indistinguishable in overall quality. Where the two frontrunners diverged was economics and speed. Generating 1,000 replies cost an estimated $1.90 with North, compared with just $0.74 with Vale—a 61% cost reduction for the same caliber of response. Vale also generated answers 18% faster, an advantage that compounds quickly in live queues where wait times shape customer satisfaction. ## How the Benchmark Worked To avoid the pitfalls of automated grading, researchers built the evaluation around human judgment. They drew 200 anonymized customer-support tickets from a software company and asked each of the five models to produce a single proposed reply per ticket. Three experienced support managers then graded every response blindly, unaware of which model wrote which draft. Scores blended three criteria that matter on the front lines: factual correctness, usefulness, and tone. ## Same Top Score, Different Strengths—and a Steep Drop Below Even though North and Vale tied overall, they did not excel in the exact same situations. North performed best on billing disputes, where careful framing and policy nuance often matter most. Vale, by contrast, led the pack in technical troubleshooting. For organizations handling large volumes of both inquiry types, the results hint that routing tickets by category—rather than relying on a single flagship model—could yield better outcomes than chasing a marginal overall score. Looking beyond the top two models reveals a real trade-off between budget and quality. The lower-cost tier—Ember, Kite, and Harbor—pushes expenses down to between $0.31 and $0.08 per 1,000 tickets. Yet that savings comes with a clear slide in performance, as average scores fall into the 70s and bottom out at 71.0 for Harbor. While a budget model might still help with low-stakes triage or internal summaries, relying on the cheapest tier for customer-facing replies risks introducing factual errors or unhelpful phrasing that human agents will have to rewrite anyway. ## Where Every Model Fell Short Before support leaders rush to deploy Vale across their entire help desk, the benchmark highlights several crucial blind spots. First, all five models performed poorly whenever a ticket depended on an image attachment. In software support—where customers routinely send screenshots of error messages or account screens—a model that struggles with visual context will fall short, regardless of how well it writes plain text. The test environment itself also narrows what these scores can tell us. Researchers evaluated only English-language tickets from a single software company, meaning the findings may not translate cleanly to other industries or multilingual support desks. Just as important, each model was graded on a single static draft. The researchers did not evaluate tool use or agentic actions—such as looking up an account or issuing a refund. Drafting a polite, accurate paragraph is very different from safely taking action inside live company systems. ## The Takeaway for Support Teams For teams looking to draft English text replies in a software support workflow, the clearest lesson is to stop paying a premium for negligible gains at the top of the leaderboard. Vale matches North’s quality at less than half the cost and with meaningfully better speed. Still, choosing a model on draft quality alone is only a starting point. Until teams test how models handle screenshots, connect with internal tools, and serve non-English customers, human oversight remains essential—especially when a ticket’s key clue lives inside an attached image.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Hydrogen: # Raincheck Raises $8 Million for Municipal Flood-Risk Software *The Austin-based company combines drainage maps and rainfall forecasts to help cities prioritize maintenance before storms hit.* Raincheck, an Austin-based developer of flood-risk software for local governments, announced today that it has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing the company’s total funding to $10.5 million. Using an approach detailed in [Raincheck’s methodology](https://example.com/raincheck-methodology), the company maps drainage systems and combines those maps with rainfall forecasts so cities can decide where to send maintenance crews before a storm. Its forecasts serve as recommendations, with emergency managers retaining control over deployments and public warnings. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana, though the company declined to identify them because some contracts have not been publicly announced. Competing alongside flood-planning vendors such as HydroMap and municipal engineering contractors, Raincheck says it will use the new funding to hire eight engineers and expand into Florida.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million to help cities plan for floods *The Austin company pairs drainage-system maps with rainfall forecasts so cities know where to send maintenance crews before a storm arrives. Basin Ventures led the seed round.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. The funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." The company has published [an explanation of its forecasting methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Google: Gemini 4 Hydrogen: # QueueZero Launches AI Meeting Assistant Proxy With Live Lesson in Literalism QueueZero launched an artificial intelligence meeting assistant called Proxy on Wednesday, introducing a tool designed to attend meetings, produce notes, and assign follow-up work to software agents. The software enters public beta next week at a price of $30 per user per month. While the product is aimed at streamlining post-meeting workflows for enterprise teams, an unscripted exchange during its launch demonstration highlighted the complexities of automated task delegation in live business settings. Proxy is built to listen to workplace discussions and translate them into structured outcomes. Rather than serving solely as a passive note-taker, the assistant monitors conversations to identify next steps and route assignments. During Wednesday’s demonstration, a customer in the audience inadvertently tested the system’s boundaries by saying, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an operational instruction. Moments later, the assistant created two distinct tasks and assigned both to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed at the output, completed the second task by delivering the pricing details, and deleted the first. While the assistant failed to register the conversational nuance, it proved remarkably attentive to customer feedback. Beyond the onstage moment, the incident underscored a broader challenge facing developers of autonomous workplace tools: distinguishing casual remarks from formal directives without slowing down automated workflows. As organizations evaluate AI agents capable of acting on spoken dialogue, clear boundaries around permissions and human oversight remain a priority for enterprise software buyers. To address such concerns, QueueZero says Proxy incorporates specific safeguards around outreach. According to the company, the platform requires confirmation before an external message is sent. Internal tasks, by contrast, can be created automatically, which is the setting that placed the two items on Wu’s list. For organizations that prefer stricter oversight over how meeting dialogue is converted into action items, QueueZero stated that administrators can disable automatic task creation altogether. With Proxy entering public beta next week, QueueZero is betting that organizations are ready to let software handle routine meeting follow-ups, provided administrators and users retain clear controls over how automated tasks are generated and approved.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Hydrogen: # HelioDock Unveils Remote Fleet Software for Warehouse Robots Alongside $24 Million Series B *The autonomous robotics company will charge $90 per robot per month for Night Shift starting September 15, after a 14-month private beta marked by fewer reported stoppages and rural LTE limits.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, introducing software that enables operators to monitor robot fleets remotely from outside the warehouse while announcing a $24 million Series B funding round led by North Coast Ventures. Scheduled to reach general availability on September 15, Night Shift moves HelioDock's remote fleet-management software out of a 14-month private beta and into commercial release at $90 per robot per month. The new Series B financing brings the robotics company's total funding to $39 million as it enters a fleet-management software category where Relay Robotics and YardOS already offer competing products, and where early customer testing showed that remote oversight depends on reliable network connections. Autonomous warehouse robots can encounter unexpected obstacles on a facility floor, leading to unplanned stoppages that interrupt operations. Without software that lets staff monitor and manage fleets from off-site locations, resolving an isolated stall can require warehouse personnel to travel back to a building outside regular hours. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said in explaining the operational issue the software targets. Night Shift is designed to give logistics teams remote visibility into their robot fleets so operators do not have to handle every disruption in person at the warehouse. Ahead of Thursday's public launch, HelioDock spent 14 months testing the software in a private beta alongside commercial logistics operators. Nine logistics companies participated in that beta trial, placing a combined 310 autonomous robots under Night Shift's remote monitoring across their warehouse facilities. According to HelioDock, beta customers experienced 17% fewer unplanned robot stoppages while using the remote-monitoring platform during those 14 months. That 17% stoppage reduction is a company-supplied figure and has not been independently verified. Because no independent verification accompanied Thursday's announcement, prospective logistics buyers evaluating the platform only have HelioDock's internal beta numbers as a measure of how much unplanned downtime the software prevents. Customer feedback from Fulcrum Logistics illustrates how Night Shift's practical utility shifts with local network quality. Aaron Pike of Fulcrum Logistics, who participated in the private beta, said the software worked well inside facilities backed by strong connectivity. During rural LTE outages, Pike said, Night Shift became less useful for overseeing warehouse operations. That drop in utility reflects how HelioDock's autonomous robots are programmed to respond when facility networks or cellular links fail. Whenever connectivity drops, remote control through Night Shift becomes unavailable, preventing off-site operators from stepping in to direct affected machines over the network. Under those conditions, the robots return to their locally configured behavior, operating according to on-site settings rather than remote direction. Thursday's software unveiling arrived alongside new venture financing for HelioDock's autonomous robotics and software business. North Coast Ventures led the $24 million Series B investment round, lifting the company's total funding to $39 million as it prepares to distribute Night Shift commercially alongside its existing autonomous warehouse robots. When general availability opens on September 15, customers deploying Night Shift will pay a recurring fee of $90 per robot per month for access to the remote-monitoring platform. That per-robot pricing moves HelioDock out of its nine-company private beta and into direct software competition with Relay Robotics and YardOS, two vendors that already have competing fleet-management products available to warehouse operators.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Hydrogen: # Startup Spotlight: LedgerLark Brings AI Agents to Post-Trade Operations Most financial technology pitches focus on the moment a trade is placed. LedgerLark is focused on what happens afterward, when records do not quite line up. The startup builds AI agents for post-trade operations at small investment firms, tackling the manual reconciliation work that sits behind the scenes of asset management. Co-founders Mina Batra and Owen Carr came to the problem from complementary sides of the back office. Batra previously spent seven years in operations at a regional brokerage, while Carr built reconciliation systems at a large bank. Batra began exploring the idea after watching a six-person team spend three days resolving a single trade mismatch caused by a copied account number. It was not a dramatic market failure, just a routine clerical error whose ripple effects consumed dozens of staff hours. For lean investment firms without large operations departments, those delays pull staff away from other work. ## Automation With Guardrails LedgerLark’s software is designed to absorb that repetitive detective work without taking humans out of the loop. Its AI agents compare confirmations, custodian records, and internal ledgers to spot discrepancies and identify likely causes. Rather than acting autonomously, the agents propose corrections for human approval. Crucially, the product does not execute trades or move money. That boundary reflects a pragmatic reality in financial operations, where auditability and control often matter more than raw speed. For small firms wary of giving automated systems direct authority over client assets, limiting the software to investigation and recommendation may make adoption easier to consider. Still, financial data is messy, and reconciliation errors often stem from edge cases across counterparties. How consistently AI agents perform across varied workflows remains an ongoing test. ## Early Traction and Seed Funding Even at an early stage, LedgerLark has found initial commercial footing. Five firms are now paying customers, with annual contracts ranging from $36,000 to $110,000. According to the company, one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the software. While a single reported outcome does not guarantee similar gains for every firm, it suggests the tool is addressing a real operational bottleneck for its early users. To expand its engineering and onboarding efforts, LedgerLark recently raised a $3.5 million seed round led by Motive Row. That funding gives the team room to refine the product, though turning early deployments into repeatable growth will take time. ## Competition and Long Sales Cycles LedgerLark enters a market where buyers already have options. Its competitors include established reconciliation vendors as well as internal bank software that many firms rely on, even if those tools still leave gaps filled by manual checks. Convincing small investment managers to switch or add a new layer to their stack requires proving both reliability and smooth integration with existing systems. Sales timelines present another hurdle. Financial institutions, even smaller ones, may require lengthy security and procurement reviews before granting third-party software access to sensitive trade and ledger data. Those evaluations can stretch out revenue timelines and consume significant attention from a young team. It also remains an open question whether LedgerLark’s sharpest advantage stays with smaller firms lacking internal engineering teams, or if it can eventually serve larger institutions. For now, Batra and Carr are focusing on one of finance’s least glamorous corners. If LedgerLark can earn trust through steady accuracy and navigate slow procurement cycles, it has a plausible path to making post-trade cleanup a little less manual.

Z.ai: GLM 5.3 Prime: # LedgerLark Takes On the Post-Trade Exception Queue Mina Batra spent seven years in operations at a regional brokerage, and one incident from that stretch stuck with her: a six-person team spent three days resolving a single trade mismatch, caused by an account number someone had copied wrong. The trade was fine; the paperwork wasn't. Untangling it consumed roughly eighteen person-days of effort. Batra began exploring what became LedgerLark after that episode. She and co-founder Owen Carr, who built reconciliation systems at a large bank, now run a startup building AI agents for post-trade operations at small investment firms. The company has raised a $3.5 million seed round led by Motive Row. ## The insight Batra's takeaway from those years wasn't that trades are hard; it's that the exception queue — the daily pile of mismatches between what was traded, what was confirmed and what the custodian says happened — is where small firms quietly lose time. Firms of this size are often too small for dedicated reconciliation engineering teams and too large for the work to stay trivial. LedgerLark's bet is that this middle band is underserved. ## What the agents do LedgerLark's software compares trade confirmations, custodian records and internal ledgers to surface mismatches. When it finds one, it proposes a correction for a human to approve or reject. The product does not execute trades and does not move money. Those boundaries matter. They keep LedgerLark out of the highest-risk parts of the workflow, and they also cap how far the software can reach. A related open question: how often the agents propose the wrong correction, and how much review each proposal actually needs. If humans must vet everything carefully, the value is time saved, not headcount avoided — and that's a harder, slower sale. ## Early traction, with caveats Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure is self-reported and from a single account, so it's suggestive rather than conclusive. It does point at the right metric, though: for an operations tool, what matters is how quickly exceptions clear and how often corrections get reopened — numbers a prospective buyer will want to audit. Five customers is also a small base from which to judge repeatability. Small investment firms differ in their ledgers, custodians and internal systems, and whether LedgerLark's agents can be deployed across that variety without heavy customization will determine whether this is a software business or a services business in a software costume. ## Competition and sales cycles LedgerLark isn't entering empty territory. Established reconciliation vendors already sell into this space, and bank-built internal software, while not marketed, shows what's technically achievable. If incumbents bolt AI features onto existing products — a plausible response given the attention agents are getting — LedgerLark's differentiation will have to come from being built for smaller firms rather than retrofitted for them. Then there's the mundane obstacle. Financial institutions tend to run lengthy security and procurement reviews before adopting new software, and post-trade tooling sits close enough to regulated processes that those reviews are unlikely to be light. Small firms may move faster than large ones, but "faster" can still mean quarters — and a seed-stage company needs a pipeline built to survive that drag. ## What to watch The founding insight is narrow in a good way: a specific, well-documented pain rather than "operations" in general. The guardrails are sensible. The early numbers are interesting, if thin. The questions that will decide the outcome are unglamorous: Are the proposed corrections accurate enough that reviewers genuinely save time? Do deployments generalize across firms? Do sales cycles fit the runway of a company that just raised $3.5 million? None of that is settled. But the underlying problem is real, tedious and measured in three-day incidents — which is exactly the kind of problem worth a closer look.

Google: Gemini 4 Hydrogen: C

Z.ai: GLM 5.3 Prime: C

Google: Gemini 4 Hydrogen: # Cloudnote Left Customer Prompt Excerpts Accessible Without Authentication for Four Hours *Logs show snippets from 1,842 user sessions on a debugging page on Monday, though Cloudnote says it has found no evidence that anyone besides a reporting researcher viewed it.* Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication for four hours on Monday, with logs showing snippets from 1,842 user sessions during that window. Each excerpt could include up to 200 characters from a user’s prompt. Cloudnote, which has 3.2 million registered users, took the page offline 19 minutes after a researcher notified the company of the issue. Full documents, passwords and payment information were stored elsewhere and were not present on the debugging page. Although the incident touched a small fraction of Cloudnote’s user base and excluded complete files, even brief prompt fragments carry meaningful privacy risks. Up to 200 characters of text can still capture sensitive questions, personal details or proprietary work context that users expect to keep private. Important unknowns remain as the company continues its review. Cloudnote says it has found no evidence so far that anyone besides the notifying researcher accessed the page while it was exposed. Because its investigation is not yet complete, however, it remains unconfirmed whether any other visitors viewed the excerpts or how the debugging page became accessible without authentication.

Z.ai: GLM 5.3 Prime: # Cloudnote Says Customer Prompt Excerpts Were Exposed on Unprotected Page for Four Hours *A debugging page displayed up to 200 characters from 1,842 user sessions before Cloudnote removed it 19 minutes after a researcher's alert. The company says no one else accessed the page, but its investigation isn't finished.* A debugging page that displayed excerpts of customer prompts was accessible online without authentication for four hours on Monday, a lapse that exposed snippets from 1,842 user sessions of Cloudnote's AI note-taking service. The excerpts could include up to 200 characters from a user's prompt. Cloudnote removed the page 19 minutes after a security researcher notified the company of the problem. The company says it has found no evidence that anyone besides the researcher accessed the page, but its investigation has not been completed. The exposure's scope has limits: full documents, passwords and payment information were stored in separate systems and were not present on the page, Cloudnote said. But even short prompt excerpts can surface details users considered private, and the incident comes as the service counts 3.2 million registered users — far more than the 1,842 sessions affected. For now, the company's assurance that the researcher was the page's only visitor rests on a review it has not finished. How the debugging page came to be reachable without a login, and why the lapse went undetected until a researcher found it, are questions the investigation is expected to address.

Google: Gemini 4 Hydrogen: 6

Z.ai: GLM 5.3 Prime: 6

Google: Gemini 4 Hydrogen: # Arcwell Launches AI Customer Support Platform Relay Arcwell has announced Relay, a new artificial intelligence platform designed to help customer service teams by examining support conversations and suggesting answers to agents. Priced at $65 per agent each month, the software is built to assist representatives across several everyday workflows. Alongside recommending responses as agents work through incoming inquiries, Relay can create conversation summaries and identify customers who may be preparing to cancel. Arcwell CEO Priya Shah framed the platform around cutting down wasted effort inside support departments. “Support agents spend too much time searching for information that their companies already have,” Shah said. To evaluate the software before its broader rollout, Arcwell tested Relay with 12 companies over a six-month period. According to the company, agents using the platform during that trial closed support tickets 14% faster. Even with those speed gains, the figures released alongside the announcement leave one part of Relay’s pitch unverified. While Arcwell highlights the platform’s ability to spot customers at risk of canceling, the company did not release customer-retention results from the six-month test, leaving open whether those early warnings help businesses keep more customers. Adopting the platform also carries important prerequisites around sensitive company records. Because Relay relies on internal context to generate suggested responses and analyze customer exchanges, businesses deploying the software must grant it access to both their support transcripts and internal help documents. Arcwell says it has put safeguards in place to protect that material from broader exposure. According to the company, customer data processed by Relay is encrypted and is not used to train shared models. Organizations adopting Relay can also limit how long their information remains stored, choosing between a 30-day or 90-day retention period. For prospective buyers, evaluating Relay comes down to balancing its $65-per-agent monthly cost and reported 14% improvement in ticket-closure speed against the internal data access it requires—while waiting for evidence that its cancellation warnings deliver measurable retention gains.

Z.ai: GLM 5.3 Prime: #