Solutions
The same scan answers several different questions depending on who is asking. Below: by what you need, by the framework you report against, and by the regulatory context you work in.
By need
01
“How many agents do we actually have?”
Teams ship agents faster than security can inventory them. Amajoni walks your AWS account and returns the list — including the ones nobody registered. Every finding is backed by the evidence that produced it, so you can tell a real agent from a scheduled job with an ambitious name.
02
“Which of these could actually hurt us?”
An inventory without severity is just a longer to-do list. Each agent is scored 0–100 against the IAM permissions it holds, with the specific permissions driving that score named individually — so remediation starts with a concrete line to remove, not a category to investigate.
03
“Can we produce the report before we're asked for it?”
Findings map to six frameworks and are stated Met or Not Met against the scan that produced them. Not Met is reported as Not Met. The value of an evidence pack is that someone else can check it.
04
“Who is accountable for this one?”
Every agent carries an owner, an approval state and a review cadence. Remediation usually stalls on nobody knowing who is allowed to switch an agent off — this is the record that answers it.
By framework
Amajoni maps findings to the controls below. It does not certify you against any of them — no tool can. It produces the evidence, and says plainly where a control is not met.
Inventory and risk-mapping controls — whether AI systems are identified, categorised and assigned owners.
Record-keeping and human-oversight obligations: whether an inventory exists, whether high-risk systems are governed, and whether someone is accountable for each.
Which agents can reach personal information, and whether their access is proportionate to their purpose.
Logical access and change-management evidence for AI workloads — over-privileged roles, approval state, review recency.
Prompt-injection pathways, excessive agency and insecure plugin/tool access, detected from real permissions and exposure.
Technique mapping for what a compromised agent could do next — credential access, lateral movement, exfiltration.
By industry
One of these is a live pilot and is marked as such. The rest describe the regulatory pressure we built for — not customers we have deployed to. We will add case studies when we have them.
Model risk management already requires an inventory of models in use and an owner for each. Agents with production data access are inside that perimeter whether or not they were registered as models.
Student records carry strict handling duties, and agents built to answer questions about students frequently hold far broader data access than the task requires.
Early access — Brown Hat Academy pilotPublic-sector AI adoption carries residency and accountability requirements that most AI tooling, built for US and EU regimes, handles poorly.
If you process customer data, your customers will eventually ask what your AI agents can reach. Being able to answer with a scan rather than a spreadsheet shortens the review.
Thirty minutes, a read-only connection to one account, and you leave with your actual agent inventory — not a slide about ours.