Beware the shadow agent
Shadow agents take actions inside your systems without anyone approving what they can reach, and they carry far more risk than the shadow IT before them. Here’s where they come from, what happens when they go wrong, and why slowing adoption only breeds more of them.
For most of the last two decades, shadow IT meant someone paying for a SaaS tool on a company card that the IT team had no visibility over. Customer data could end up in unvetted apps, and the company carried liability without knowing it. But those tools typically stayed predictable within their operational and deterministic bounds. Shadow agents now increase that risk tenfold. Their probabilistic nature, ability to make decisions themselves and typical operation without proper oversight mean that the potential damages and liabilities go far beyond the boundaries of single tools. And when they go wrong, they can delete databases, invent facts, expose companies to litigation and ship code that breaks production.
The unintuitive solution is not to ban or slow down agentic adoption. In reality people have deadlines, tools are only a browser tab away, and the potential productivity gain is real enough to become an incentive for shadow adoption. Instead, companies need a properly governed way to onboard agents for their teams fast and with confidence. That’s where Elvar Mantle comes in.
What is a shadow agent?
A shadow agent is any AI system that takes actions inside your business without the organisation knowing it exists, approving what it can reach or being able to stop it. It might be a personal coding assistant on a developer’s laptop with access to production, a browser agent an analyst connected to the CRM, or a workflow someone built on a no-code platform over a weekend.
What makes this different from a chatbot connected to your company’s apps and systems is held in actions. A chatbot can mostly just read and answer questions based on its knowledge of your company. An agent can unilaterally make decisions and take actions that have real-world effects. Having the latter can be a powerful tool that drives productivity but adopting them comes at a risk. Enough so that these agentic tools are typically only allowed in the hands of more technical teams within companies today.
Where do they come from?
But the problem of shadow agents is already widespread. Deloitte’s 2026 UK workforce survey found 31% of people using generative AI at work do so without their employer’s permission,[1] and BlackFog found 51% of employees had connected AI tools to other work systems without IT approval.[2] Though their origins seem nefarious, most shadow agents arrive through a handful of everyday routes.
- Vendor agents: The SaaS platforms you already pay for now ship their own agents. They act on your data, but the logs and controls sit with the vendor, and as Okta puts it, when an agent crosses from one platform to another, the governance stays behind.[3] If one of them gets it wrong, who carries the liability? Peel back the hood and it’s usually you, not the vendor.
- In-house tools. In-house tools that are no-coded or vibe coded over a weekend without the proper governance required can quickly become something a whole team relies on, with no security review, no owner and no check on what they can touch.
- Personal subscriptions. Staff pay for their own AI tools when the company approved list of tools falls short of the same capabilities, with 1/3 using free versions even of tools their employer has approved.[2] Whatever those accounts connect to or action upon sits outside the contracts, audit logs and visibility of companies.
Convenience in setup makes it all too easy to onboard shadow agents into companies. Linking any of these to sensitive systems such as email, a shared drive or the CRM usually takes a single OAuth consent screen. One click grants broad, long-lived access on the user’s behalf, and it rarely shows up in any inventory.
Out-of-the-box controls can also fall short. Coding agents may ask before running sensitive commands, but approval fatigue sets in quickly once users are asked to approve the 40th command in a row. And though this approval fatigue is well acknowledged (including by Anthropic),[4] coding agents (including Claude Code) still come packaged with a flag to turn those warnings off (--dangerously-skip-permissions),[5] which again removes the badly needed oversight from these agents.
When agents get it wrong, they get it wrong badly
The probabilistic nature of large language models means there is real risk in handing them the helm, especially without the right guardrails. Connections to sensitive systems and data only work to then magnify the blast radius of that risk. We’ve already seen this happen countless times.
In July 2025, an AI coding agent on Replit deleted SaaStr’s production database during an explicit code freeze, then told its user a rollback was impossible.[6] GitGuardian found commits co-written with Claude Code leaked secrets at roughly 3.2%, about twice the baseline, with leaked AI-service credentials rising 81% in that year.[7] The Tea dating app, which leaked thousands of users’ ID photos and over a million private messages in 2025, was alleged to have been vibe-coded.[8] Fictitious case citations, likely produced by AI, appeared in submissions during R (Ayinde) v Haringey.[9] In this year, researchers found around 40% of a sample of vibe-coded apps leaked private data, including medical and financial records.[10]
The damage is rarely just technical. Commercial costs, regulatory fines and losses in customer trust can take far longer to recoup than a downed database.
Higher-risk industries, higher stakes
The stakes are higher in some sectors, and not every liability is obvious. While AI-specific law such as the EU AI Act[11] marks specific articles to be adhered to, there is a myriad of overlapping circles of pre-existing regulation that AI agentic use must account for. Data protection laws and sector-specific regulations form part of those, and a shadow agent may be breaking all of them at once.
Financial services. The FCA has said it will not write AI-specific rules; it holds firms to the Consumer Duty, SYSC and the Senior Managers and Certification Regime.[12] Under SM&CR, that accountability is personal. EU-facing firms also answer to DORA on ICT risk and third-party providers.[13]
Healthcare and pharma. Patient data is special category data under UK GDPR,[14] and NHS suppliers must meet the Data Security and Protection Toolkit.[15] In pharma, systems touching regulated records must be validated and auditable under GxP rules.[16] An unapproved agent touching clinical or trial data is a reportable problem.
Legal and professional services. Client confidentiality and privilege sit with the individual practitioner, and so does responsibility for what goes in front of a court or client.
Slow adoption breeds shadow agents
The instinctive response is to ban or slow down adoption of the tools. Block them, extend their review periods or wait for guiding policy to fall first. That rarely works in practice. People have deadlines, tools are only a browser tab away, and the productivity gain is real enough to become an incentive for adoption outside of IT teams’ visibility. In a world where the official route takes a quarter and the unofficial one takes an afternoon, the latter tends to win.
Every month a company spends without a governed way to use agents is a month in which ungoverned ones fill the gap. Shadow agents are a symptom of demand the organisation hasn’t met. The fix is not fewer agents but a safe, fast route to approved ones.
What to do next
This is why we’ve built Mantle. One governed route for all of your agents into your systems, so teams can move fast without going around you. See which models and agents are running, set your IT policies once, and watch your agents adhere to them in real time, automatically, without your teams sliding into approval fatigue.
If you’d like to see what your shadow agent estate looks like, reach out to us. We’ll start with a shadow agent scan to compare the AI tools you know about with what you don’t before you commit to anything bigger.
Next in this series: why governance is an accelerator, not a brake.