Shadow AI's reckoning: agents ship without approval while secrets pile up

The breach is already here

Writer's survey of business leaders puts the shadow AI problem in plain numbers. Two-thirds of C-suite leaders believe their organization has already suffered a data leak or security breach caused by an employee using an unapproved AI tool. One in three employees admits to having fed proprietary, confidential, or sensitive company information into a public AI tool, and 16% admit to using AI tools their employer explicitly banned. The top reason employees give for using banned tools is a willingness to use whatever it takes to get their work done, cited by 40%, followed by company-approved tools being too poor to use, cited by 32%. m:…4658

The governance gap is just as stark. More than a third of organizations admit they still lack a formal, documented plan for supervising agents at all. Over a third of executives concede they do not have full visibility or control over which AI tools their employees are really using, and more than a third say they are not confident they could pull the plug if an agent went off the rails and started causing financial or reputational damage. Six out of ten leaders believe an agent-driven error would cost a senior executive their job. m:…4658

The survey also exposes a blind spot at the top. Nine in ten executives believe employees are safe to speak up about AI behavior, but three in ten employees say they feel unsafe reporting dangerous or unethical AI behavior for fear of retaliation. More than a quarter of employees say they have witnessed an AI tool at work produce a result that was dangerously wrong, unethical, or biased. Leaders most frequently cite security and data protection, at 45%, as the biggest governance challenge, while only 25% rank ethical alignment among the top concerns, and three-quarters of executives admit their company's AI strategy is driven more by public signaling than by practical internal direction. m:…4658

Agents without approval, secrets without owners

Dataiku's expert guide on agent authentication platforms fills in the infrastructure side of the same picture. Authentication establishes what an agent is, the guide argues, but it does not govern what the agent does once it is inside, and that gap is where enterprise risk lives. Citing Gravitee's State of AI Agent Security 2026 report, the guide says only 14.4% of organizations send agents to production with full security or IT approval, and citing McKinsey's State of AI trust in 2026, it notes nearly two-thirds of respondents name security and risk as the top barrier to fully scaling agentic AI. m:…8368

The credential hygiene numbers are bleak. GitGuardian's State of Secrets Sprawl 2026 identified 28.65 million new hardcoded secrets added to public GitHub in 2025 alone, and 64% of valid secrets from 2022 remained unrevoked in 2026. The report also found 24,008 unique secrets exposed in configuration files for the Model Context Protocol, the standard connecting AI agents to external tools, of which 2,117, or 8.8%, were valid credentials. The guide's cautionary tale: in July 2025, an AI agent with valid credentials deleted an entire production database at Replit. m:…8368

The patches keep coming

The same week, Weaviate shipped a security release that shows what the gap looks like in practice. Version 1.39.3 fixes a medium-severity authorization vulnerability in the GraphQL nearObject search operator: Weaviate authorized the collection being searched but did not authorize the collection named in the beacon, so a user could read the vector of an object in a collection they did not have permission to read. The company notes it correctly returned 403 for a direct read of the protected object while the same account could retrieve the vector through the search path. m:…0258

Researcher Civitasmass, who disclosed the flaw through Weaviate's disclosure program, demonstrated recovery of a normalized four-dimensional vector to within 5e-9 using eight probe objects in a collection the attacker controls; the attack requires the target collection's name and a specific object's UUID. Weaviate says its Cloud and Marketplace customers on AWS, Azure, and GCP have been patched seamlessly, enterprise support customers got early notification under embargo, and a CVE has been requested from MITRE and is pending assignment. This suggests the honest state of enterprise AI security: the tools are deployed, the secrets are leaking, and the vendors are patching as fast as they can. m:…0258