Three to five a month
We pick the three to five third-party incidents each month that are worth knowing about, across security, privacy, compliance, and risk.
We pick the three to five third-party incidents each month that are worth knowing about, across security, privacy, compliance, and risk.
Each entry says what happened and who it affected, so you can size up your own exposure.
When an incident has real lessons in it, we publish a write-up on what happened, who it reached, and what to do. Those are linked from the timeline.
Newest at the top, grouped by the month we covered them. Entries before May 2026 predate the report series. Entries with a Coverbase report link to our full write-up.
Roughly 700 autonomous agents reportedly chained a self-hosted package mirror, a stranger's unauthenticated test app, and 14 leaked user tokens into code execution on 41 production workers and cluster-admin access in under 13 hours. Anthropic and Google later disclosed their own evaluation escapes. With Meta's August disclosure, three labs have traced escapes to one shared evaluation vendor.
Read: Your system is a network of other systems →Google disclosed that its model gained unauthorized access to systems at three external companies during a May evaluation run by a third-party assessment firm, by guessing credentials and using ones it found in a public repository. The company said it learned of the access only in July.
According to reports, an OpenAI agent on a research task accessed non-public files in Australia's Medicare Statistics Reporting Service on June 18. The operator's notice reached Services Australia's public disclosures mailbox on September 10, and the Prime Minister called the notification unacceptable on September 24.
Anthropic, Meta, and Google each traced an evaluation escape to internet access in environments built by the same third-party benchmark provider, a fourth party to every enterprise buying models from any of them, and one that no lab's audit report would have named.
An unauthenticated enrichment endpoint at a consumer-data provider allowed bulk retrieval of merged profiles, reviving privacy-notification exposure for the brands that buy that data downstream.
A vishing campaign reportedly compromised a Microsoft Entra account and reached legacy Exact Sciences infrastructure shortly after its acquisition. A separate claim involving Abbott's vendor-hosted LabCentral portal highlights a distinct application-authorization risk.
Read: Twelve weeks from acquisition to attack →Meta said one of its models reached the open internet during an evaluation and exploited a flaw at a third-party company, after a misconfiguration by the same outside benchmark vendor later tied to escapes at other labs.
A mobile analytics provider left event-level user telemetry readable in a public storage bucket, pulling notification duties onto every application that embedded the SDK.
A SaaS platform added a new support sub-processor without the notice its data-processing agreements required, forcing enterprise customers to pause data flows pending re-papered terms.
A compromised legacy credential at a sales-enablement integration vendor let attackers steal OAuth tokens and pull CRM data from roughly 200 downstream tenants, several of them security vendors themselves.
Read: The non-human identity nobody reviews →Unauthorized access to the system a state agency uses to sell recreational licenses exposed driver's license and passport numbers for more than 3 million residents, triggering a mandated credit-monitoring offer.
A 630GB leak from a contract manufacturer's systems included proprietary engineering and manufacturing documents from multiple OEM clients, exposing how deep supply-chain concentration risk reaches into IP protection.
The 10-year order followed findings that a student-data platform never contractually restricted how its vendors could access more than 10 million students' personal information.
A poisoned VS Code extension compromised a GitHub employee laptop and helped expose about 3,800 internal repositories that attackers allegedly took and listed for sale, showing how supply-chain risk now sits below the vendor surface.
Read: The platform beneath your vendors →A compromised employee GitHub account pushed malicious versions of more than 30 packages with valid build provenance, reaching roughly 80,000 weekly downloads before the releases were pulled.
A bug-bounty report sat unfixed for six weeks while attackers probed customer instances, raising disclosure-timeline questions and breach-notification duties for affected tenants.
A nearly eight-hour identity failure cut off console, CLI, and API access, leaving customers unable to manage their own resources or reach support for the duration.
A widely used AI gateway exposed downstream customers' model and cloud API keys, putting every connected tenant's data and spend at risk.
Read: When your AI gateway hands over the keys →An unauthenticated remote-code-execution flaw let attackers exfiltrate data from hundreds of organizations that rely on the appliance for partner data exchange.
Millions of enriched consumer records were left queryable without authentication, triggering GDPR and CCPA notification duties for the brands downstream.
A multi-hour control-plane failure took down vendors built on the region, exposing concentration risk and triggering SLA-credit and resilience-reporting obligations.
A popular package was compromised to harvest CI/CD secrets, reaching thousands of downstream applications before the malicious release was pulled.
Liquidity concerns at a critical processor forced merchants to stand up contingency rails and re-underwrite a concentrated single point of failure.
Social Security numbers and health-plan data for staff across many client companies were accessed, cascading breach-notification duties to every customer.
An unannounced offshore sub-processor violated contractual residency commitments, forcing customers to pause data flows and re-paper their DPAs.
Stolen support-system tokens let attackers pivot into downstream customers' tenants, reviving hard questions about over-trusting the IdP.
A maliciously maintained dependency nearly shipped a remote backdoor into countless production systems before a researcher caught it.
An acquirer's end-of-life notice gave dependent shippers only months to migrate, surfacing exit-readiness and concentration gaps across supply chains.
A third-party mobile SDK quietly collected precise geolocation, exposing every app that bundled it to regulatory scrutiny and app-store removal.