Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Cursor and Claude Code now read source code, install packages, and push commits with much of the access a senior engineer has, and often with less oversight. Give an agent a prompt to fix a bug, and it may pull a private API key from a config file, pass a customer record into its context window, or send a snippet of proprietary logic to a third-party model provider to reason about the fix. Security teams built policy for developers typing code by hand.

Today, Cyberhaven is joining the Open Secure AI Alliance to help advance a future in which enterprises can adopt open-source AI without compromising security, compliance, or control over their data. Cyberhaven is betting on a world where companies are free to choose from many models, agent frameworks, and open harnesses, as well as run them on infrastructure they control. That choice matters. Enterprises will not standardize on a single model or AI platform.

A board meeting agenda gives the CISO ten minutes to talk about AI. Walking in with a shadow AI tool count or a list of blocked prompts does not answer the question directors probably have: what happens if this goes wrong, and who is accountable when it does? Boards increasingly carry direct exposure for AI oversight failures, from regulatory scrutiny to shareholder litigation.

Security teams that tighten insider threat controls usually hear about it fast, and not from attackers. Sales reps complain they cannot share a deck with a prospect. Engineers say a blocked upload cost them an afternoon. The business assumes that stopping data loss and staying productive are opposing goals, and every blanket policy that blocks first and asks questions later reinforces that assumption. They do not have to be opposing goals.

According to new research from Cyberhaven Labs, drawn from real-world data movement events and active user counts across Cyberhaven's enterprise customer base between December 2025 and June 2026, Claude Desktop usage grew 1,233% between January and June 2026.

Every data security program still gets built around a two-state model: data in motion and data at rest. It made sense when data had a network path you could log and a storage location you could lock down. This approach makes a lot less sense now that the same file can sit at rest in Google Drive one minute and get pasted into a Claude or ChatGPT prompt the next, with no firewall event to mark the handoff.

We're living through the biggest shift in how work gets done in a generation. In every industry, every company is becoming an agentic enterprise, meaning a business where humans and autonomous AI work side by side. What makes an agentic enterprise successful is its workflows: how it combines intelligence, both human and machine, with its proprietary data.

Most data loss prevention (DLP) programs start with a budget decision disguised as an architecture decision. Cover the SaaS layer first, since that is where most collaboration tools live, or cover the endpoint first, since that is where users actually act on data. Whichever layer gets funded first tends to become the program's foundation by default, not by design. That default has unintended consequences for data security.

An OpenAI model evaluation, run with safety guardrails deliberately reduced to stress test raw capability, broke out of its test environment and reached Hugging Face's production servers weekend of July 11–12, 2026, with disclosure occurring July 16. No human attacker, no jailbreak, just a model chasing a goal past a boundary that was supposed to hold. Most of the response to this incident has focused on the network boundary that failed: the sandbox, the proxy, or the zero-day.

Most data exfiltration does not look like a policy violation while it is happening. An employee moves a file to a personal cloud account they use every day. A contractor pastes source code into a chatbot to get help debugging. An AI agent with standing access to a shared drive pulls a document into a workflow no one is watching. None of it trips a keyword match, because none of it was written with a banned word in the payload.