The autonomous agents are already inside the enterprise. On Wednesday, the money moved to watch them.

In a single day, roughly $180 million in fresh venture funding — and, the day before, a $1 billion acquisition — landed on one narrow problem: how to secure, govern and verify AI systems that now act on their own inside company networks. It is the quiet half of the AI build-out — not the models, but the controls around them — and after last week's breach of Hugging Face by an autonomous attacker agent, investors are pricing it as urgent.

Securing the non-human workforce

The largest single move was an acquisition. On 28 July, data-security firm Cyera agreed to buy Oasis Security for $1 billion, folding Oasis's identity and access-management platform into Cyera's data-governance stack. The stated rationale is agents: as software identities that can read data and take actions multiply, someone has to decide what each one is allowed to touch, and who authorised it.

The venture rounds told the same story. Onyx Security — which calls itself "the secure AI control plane for the enterprise" — raised a $113 million Series B led by Bessemer Venture Partners, taking its total to $153 million barely two years after founding and four months after leaving stealth. Onyx says it monitors "each step of agent reasoning" to catch and correct unintended or malicious actions; in June, Anthropic announced an integration with the company to help enterprises adopt its models safely.

Hush Security, which describes its focus as "securing the non-human workforce," closed a $30 million Series A, lifting its total to $41 million less than a year out of stealth. And Empirical Security announced a $25 million Series A led by Brightmind Partners, aimed at predicting where AI-accelerated attackers will strike first.

The common thread is a phrase the industry has settled on almost overnight: the "non-human workforce." The agents are the new employees, and nobody gave them badges.

The other half of the trust problem: telling machine from human

If one front is controlling what agents do, the other is detecting what they make. Also on 29 July, Pangram Labs raised $9 million led by Menlo Ventures, with Haystack, ScOp, Script Capital and Cadenza participating, and launched Pangram 4 — a text detector the company says is "over 99% accurate" at spotting AI-assisted and mixed human-AI writing — alongside Pangram Image, an image detector currently in research preview.

Those accuracy figures deserve the usual caution. Detection is an arms race, not a solved problem: TechCrunch's own testing found Pangram 4 "not perfect," flagging some human-written sentences as AI-assisted, and founder Max Spero concedes that roughly one in 10,000 human documents get mislabelled. On a document, that is a rounding error; across a university's essay submissions or a publisher's slush pile, it is real people wrongly accused. We looked at whether AI detectors actually work earlier this month, and the honest answer has not changed: they are useful signals, never verdicts. What is new is that the market now believes the signal is worth $9 million more than it did a week ago.

Washington's deadline lands this weekend

The policy layer is moving on the same clock. Under Executive Order 14409, signed 2 June, federal agencies must stand up a voluntary framework by 1 August for frontier developers to submit their most capable models for government review up to 30 days before a model's planned release — time the White House intends to use to find security weaknesses before it ships.

That deadline arrives on Saturday, days after more than 1,100 employees of the leading labs asked Washington for tools to pace automated AI development. The through-line across all of it — the funding, the acquisition, the framework — is the same: the frontier is being built faster than anyone can govern it, and the trust layer is where the catching-up is now being funded.

The short version

• ~$180M in raises and a $1B acquisition hit AI-agent security and content-authenticity in a single day (28–29 July).
• Cyera–Oasis ($1B), Onyx ($113M), Hush ($30M) and Empirical ($25M) are all selling versions of the same thing: control over autonomous agents inside the enterprise — the "non-human workforce."
• Pangram ($9M) works the authenticity front, but its own "over 99%" claim comes with tested false positives; detectors are signals, not proof.
• 1 August: the White House's voluntary frontier-model review framework is due — the government's parallel move to see powerful models before the public does.