Exaforce ties every AI agent it discovers back to the employee whose permissions it borrows

The detection runs on data the buyer has already centralized. Exaforce matches agent and model provider activity against human identities, endpoint telemetry, file access and code context, inside the knowledge graph it builds from identity, cloud, software-as-a-service and code sources ^[5][7]. Into that graph go endpoint data, productivity suite activity and model provider audit and usage logs ^[7]. Ankur Singla, co-founder and chief executive, said existing software-as-a-service and endpoint tools were "never built for this era of AI" ^[9]. He also said security teams need agent activity landing where their identity, cloud and endpoint data already sit, without another rollout to get it there ^[10].

Discovery is the part a team will feel on Monday. It runs continuously, needs nothing installed, and turns up connected AI apps, coding agents such as Claude Code and Cursor, hosted agents including custom GPTs, Model Context Protocol servers, skills and development environment plugins ^[3]. Each one is tied back to the person and the permissions behind it, and a risk layer grades what turns up and flags excessive permissions or unsanctioned applications ^[4]. Much of that flagged list will be tools individual employees connected themselves.

The three incidents Exaforce points to do not all turn on the same weakness. In the June compromise of Klue Labs, a Canadian competitive intelligence company, customers' OAuth tokens were exposed and then used to pull data from their Salesforce environments ^[11]. The same technique reached more than 700 organizations through Salesloft's Drift chatbot in August 2025 ^[12]. Those two are stolen integration tokens. The Nx "s1ngularity" attack on the npm registry is the closer case for the product's premise: developers' own coding agents, Claude Code and Gemini CLI among them, were turned into credential hunters ^[13].

That distinction matters for how a buyer scopes this. An agent running with a staff engineer's permissions and a third-party token stolen from a chatbot vendor both produce log entries that look like normal work, and only the first is something an inventory of your own agents would have caught in advance ^[2][3].

HarbourVest Partners, Peak XV Partners, Mayfield Fund, Khosla Ventures and Seligman Ventures backed a $125 million Series B for the San Francisco company in May at a reported $725 million valuation ^[14]. Exaforce has raised $200 million since it was founded in 2023 ^[15], which leaves $75 million raised before that round ^[16]. SiliconANGLE's account does not include a price for AI Security.

Response is where the rollout gets uncomfortable, because the platform can revoke a session, deactivate a model provider key, isolate a device or end an agent's process, and each action has its own autonomy setting from analyst-approved to fully automatic ^[6]. Sort each discovered agent by two facts you can write down: the name of the human whose permissions it borrows, and what breaks if the process ends mid-task. Where both are known, fully automatic is defensible. Where the second is unknown, leave it analyst-approved, and make sure whoever holds the pager has the authority to deactivate a model provider key at 3am.