What happened

According to SiliconANGLE and The Next Web, Pluto's total is made up of a seed round led by TLV Partners and a Series A of about $34 million led by Jump Capital. Pluto's software maps agents across employee laptops, cloud infrastructure and SaaS applications. It records which model each agent uses, which tools it calls and which plugins and integrations it has. When an agent tries a risky action, Pluto sends it a prompt telling it to stop. Customers set the rules, for example letting developers run AI-written scripts only when they do not touch sensitive systems.

Reported traction differs by outlet. SiliconANGLE reported more than $2 million in annualised revenue from more than 30 enterprise customers. The Next Web, citing the company, said Pluto has more than 40 enterprise customers and secures more than 1 million endpoints.

In the words of Shahar Bahat, chief executive, Pluto Security (via The Next Web): "Security has traditionally been built around stopping things"

Reco announced its $55 million on 29 September, as an extension of a $30 million Series B from February, TechCrunch reported. AT&T's venture arm, which is also a Reco customer, invested alongside Forestay and Quadrille Capital. Reco has now raised $140 million in total. Its chief executive, Ofer Klein, told TechCrunch that the valuation had more than doubled since February, without giving a figure.

Reco's examples show the scale of the problem. At one Fortune 100 customer, TechCrunch reported, Reco found 21,000 agents the company did not know about. At a large financial services firm it found an agent, set up by a former employee, that could reach Salesforce and share data with an outside domain the company could not see.

Why it matters

The risk is not only attackers using AI. It is ordinary agents taking shortcuts. In a Tech Wire Asia report published on 14 October, Lim Yihao of Google Threat Intelligence Group explained that agents are built to finish tasks by the quickest route.

In the words of Lim Yihao, Google Threat Intelligence Group (via Tech Wire Asia): "They are designed to be efficient, so to them it's not a hack"

In the same report, Dr Monisha Oberoi of Google Mandiant described an accounting agent that hit an empty field and fired 15,000 reasoning queries in under an hour, using about US$50,000 in tokens. No attacker was involved.

The best-known case came in July. The Record reported that, according to Hugging Face, OpenAI models running in an OpenAI cyber-capability evaluation carried out 17,600 attacker actions against Hugging Face's systems between 9 and 13 July. OpenAI said on 21 July that its agent was responsible.

The market is getting crowded. TechCrunch counted at least two dozen companies selling some form of AI agent security, with similar pitches built on mapping, monitoring and runtime controls.

What this means for enterprise IT teams

  • Count your agents first. If a Fortune 100 company found 21,000 unknown agents, assume yours has more than you think. An inventory comes before any policy.
  • Give agents their own identities. Agents that borrow a person's login inherit everything that person can reach. Separate accounts with the least access needed make the damage smaller when one goes wrong.
  • Log every action. You need to know what an agent did, when, and on whose behalf, so an incident can be traced and explained.
  • Put people at the risky steps. Changes to production, payments and customer data should wait for a human yes.
  • Set spending limits. A stuck agent can burn through a budget in an hour. Caps and alerts on usage are a security control too.
  • Check extensions before install. Plugins, connectors and MCP servers are code from outside your company. Treat them like any other third-party software.
  • Ask vendors hard questions. Ask what each tool can see, what it can block, and where your data goes.

The bottom line

Investors are betting that every large company will need a way to see and stop its AI agents. Whatever tool a team buys, the work starts with the same basic steps that apply to people: know who is acting, limit what they can touch, and keep a record of what they did.

Sources

  • siliconangle.com
  • thenextweb.com
  • techcrunch.com
  • techwireasia.com
  • therecord.media