Two companies can run identical cloud security tools, receive identical findings, and still need completely different fixes.
Cloud security tools are good at what they do. A CNAPP, CSPM, or native cloud service will reliably find the public bucket, the permissive security group, and the unencrypted volume at scale.
And yet, Tamnoon’s State of Cloud Remediation 2026 report analyzed 14.86 million detections and found that critical alerts stay open for 150 days, with the open share of findings rising from 41% in 2025 to 53% in 2026.
The backlog grows because every finding needs context before anyone can act on it, and gathering that context is the slow part, as it has to happen per finding, inside one specific environment.
The teams closing this gap have stopped treating remediation as a single capability. They break it into separate questions, answer each one against their own environment, and let the answers compound. That’s the exact model Tami runs on.
Tami coordinates specialized skills. Each one answers a single question about a finding, then hands its answer to the next. This includes:
- What the asset is
- Whether the business depends on it
- Who owns it
- What gets worked first
- What breaks if it changes
- Whether the fix is safe to make
Every skill is configured against your environment, down to your database engines, your naming conventions, and your own definition of a sensitive asset. Agentic remediation built this way produces fixes shaped by one specific cloud.
See how Tami’s remediation skills work together, the questions each one answers, and how they turn a single finding into a clear path to remediation.
Cloud Security: 1,200 Different Remediation Problems
Remediation looks like one job from the outside. In reality, it splits into roughly 1,200 distinct problem clusters, each with its own fix path, required expertise, and safety considerations.
Closing a permissive security group has almost nothing in common with rotating an exposed key or patching a container image. The investigation, blast radius, and definition of done all differ.
A single general-purpose agent covering all 1,200 has to generalize. It produces advice that is broadly correct and locally useless, which is how most remediation guidance ends up ignored. Tami handles this by splitting the work into task-specific skills.
What a Remediation Skill Is
A skill is a named, specialized capability that answers one question about a finding. Tami coordinates them, and each skill’s output becomes the next skill’s input.
Anyone who has worked with AI skills in other tools will recognize the architecture. A capability lives as a self-contained unit rather than as instructions buried inside one large prompt, so it can be inspected, tuned, and improved on its own.
The useful part for a security team is what sits inside each one. Every skill exposes what it looks at, how it decides, and what it does when signals are ambiguous. Those settings hold one customer’s reality: their naming conventions, their database engines, their org structure, their definition of a sensitive asset.
Six skills run in the current build:
- Asset Classification: Labels what each resource is and how sensitive it is
- Crown Jewels: Identifies the highest-value assets to the business in the environment
- Owner Attribution: Resolves who owns a given asset
- Recommendations: Turns the alert feed into a short, ranked list of work
- Impact Analysis: Determines what would break if a fix is applied
- Remediation Confidence Indicator: Decides whether the fix is safe to make
The Six Questions Behind Every Fix
The skills run as a chain, and each one narrows the problem before handing it to the next. The order below is the order they execute.
| Step | Skill | Question it answers |
|---|---|---|
| 1 | Asset Classification | What is this thing, and how sensitive is it? |
| 2 | Crown Jewels | Does the business actually care about it? |
| 3 | Owner Attribution | Who owns it? |
| 4 | Recommendations | Out of everything open, what gets worked first? |
| 5 | Impact Analysis | If we fix it, what breaks? |
| 6 | Remediation Confidence Indicator | Is it safe to act? |
A remediation plan is generated only after the final skill returns SAFE, so every question ahead of it filters what reaches production.
How the Remediation Skills Workflow Works
Take a permissive security group on a production Postgres instance holding customer PII. Here is what each skill contributes.
Asset Classification
As the asset is discovered, Tamnoon scans its name, resource type, and tags against a keyword library.
The instance name matches terms in the database dictionary, which carries around 50 entries spanning engines and managed services from Postgres and Aurora through to ClickHouse and YugabyteDB. A tag match sets the environment to production. A connected DSPM feed adds the PII label.
The skill also normalizes asset types across sources. One tool may report a resource as “S3 Bucket” and another as “AWS S3.” Tamnoon unifies both into a single type with a storage category, so grouping and reporting work across every connected tool rather than one tool at a time.
Two behaviors keep the output clean. Ambiguous matches such as emails and UUIDs are ignored, and development, test, and sandbox assets are weighted down so non-production noise does not compete with real risk.
Handed forward: a production database asset labeled PII.
Crown Jewels
Severity ratings come from the detection tool and describe the finding. Crown Jewels describe the asset, which is a separate measure of what is actually at stake.
Tamnoon identifies Crown Jewels from environment, asset type and category, classification, and direct customer input. Production plus database plus PII clears the bar, so the instance is flagged.
Customers encode their own vocabulary through labels. A keyword like pci-scope can add a credit card classification when it appears in a tag value. A keyword like patient-records can force both a PHI label and Crown Jewel status when it appears in an asset name. A company that has tagged PCI scope the same way for eight years re-tags nothing. It tells Tamnoon what the tag means once.
Where a connected CMDB disagrees with Tamnoon’s own Crown Jewel mapping, the CMDB wins. The customer’s system of record has final say.
Handed forward: an asset with confirmed business impact.
Owner Attribution
Ownership is where remediation most often stalls, and the handoffs between security, infrastructure, and development are usually where context gets lost.
Tamnoon reads cloud tags to surface the owner, scanning keys including owner, createdby, created_by, and created by, matched as case-insensitive substrings. Every match is kept and listed. Here, an owner tag resolves to the data platform team.
The decision worth noting is what happens when no tag exists. The field stays blank. There is no inference from IAM history, deploy records, or contact lists. An incorrect owner routes work to someone who cannot act on it and breaks trust in every other field on the page, so the gap stays visible instead of being filled with a guess.
Manual overrides can be set per asset or inherited from an account or VPC.
Handed forward: a finding with a known owner.
Recommendations
Raw alert feeds are lists. Recommendations group alerts that share a single fix, then rank the groups so a team works from a short queue.
Grouping runs across team, application, compliance framework, crown jewel, and alert type. Grouping by team is recommended, so every recommendation lands with exactly one owner. Our security group finding joins other network exposure alerts belonging to the data platform team.
Ranking starts at the individual alert. Two machine learning models score every one: asset criticality, meaning how important the asset is in the context of this environment, and alert toxicity, meaning how dangerous this particular alert is to that asset. The same permissive rule on an isolated sandbox host and on an internet-reachable production database produces very different toxicity.
Each recommendation then receives a priority score between 0 and 1, built from four signals that each guard against a different way ranking fails. This includes:
- Typical alert: stops a group of one serious alert and ninety minor ones from outranking a group of consistently serious ones
- Worst alert: stops a single catastrophic finding from being buried under volume
- High-severity concentration: rewards groups where severity sits at the top rather than spread thin
- Volume: rewards groups that retire a meaningful amount of work in one fix
The weights are calibrated by Tamnoon and always sum to 1, so the score stays comparable across every recommendation in the tenant.
Handed forward: a ranked position in the queue.
Impact Analysis
This is the question that gets automation switched off at most companies. For each remediation path, Tami works through the operational consequences and produces three things:
- Concrete risks per execution path
- Conditional logic covering risks that apply only in certain circumstances
- Read-only verification commands written for the specific cloud provider
Six categories get weighed. These include outage and availability, application flow disruption, dependency breakage, identity and access disruption, data availability, and side effects on neighboring assets.
That last category covers what teams miss doing this by hand. Nobody forgets to check the resource being changed. People forget the resource three hops away that shared the security group.
Impact Analysis is read-only by design. It informs the decision and changes nothing in the environment. The verification commands are read-only as a property of the skill, not as a setting someone could switch.
Handed forward: a documented blast radius, plus commands the reviewer can run.
Remediation Confidence Indicator
The Remediation Confidence Indicator, or RCI, is the gate. It returns one of three verdicts.
| Verdict | Meaning |
|---|---|
| SAFE | Zero blast radius determined, or team safety confirmation received. Safe to remediate. |
| RISKY | Potential impact to the environment. A human reviews before anything happens. |
| AWAITING DATA | Key data needed to determine confidence in the fix is missing. |
Inputs arrive from both sides. Tami injects automation output including traffic analysis, security group analysis, exposure analysis, code repository analysis, and access control checks. The team contributes comments and the results of the Impact Analysis verification scripts.
The bias is conservative and stated openly. When the choice sits between SAFE and RISKY, the RCI returns RISKY. When it sits between RISKY and AWAITING DATA, it returns AWAITING DATA. Uncertainty always resolves toward more caution and more human involvement.
For our finding, traffic analysis shows no active connections on the exposed port and the dependency check comes back clean. The verdict is SAFE.
Handed forward: a SAFE verdict, the only input that produces a remediation plan.
Same Skills, Different Engine
Read the six configuration surfaces together, and they form a questionnaire about one company:
- Which database engines do you actually run?
- How do you name production, staging, and development?
- What does pci-scope mean inside your organization?
- Which system of record wins when it disagrees with us?
- How do you tag ownership?
- How is your work divided: by team, by application, by compliance program?
- What counts as someone on your team confirming a change is safe?
Answer those questions differently, and the same six skills produce different behavior. The environment is the configuration, which is what separates a remediation engine built for one cloud from a general set of recommendations.
The training data works the same way. Different dependencies, ownership structures, and blast radii produce different fixes, so Tami draws on millions of real, expert-supervised cloud fixes rather than scraped documentation.
What Happens After a SAFE Verdict
A remediation plan is generated only for findings the RCI rates SAFE. RISKY and AWAITING DATA findings get no plan at all. Plans are built through an agentic flow that constructs a tree per asset with the steps needed to close the underlying alerts.
Execution then follows the control that governs any production change under SOC 2, which calls for two verifiers before a change reaches production.
Tamnoon supplies the first one. A Remediation Expert reviews the plan and verifies it against the environment before it goes anywhere. Half of the control is handled for you.
The second verifier is someone on your team, and that confirmation is the same explicit comment the RCI treats as team safety confirmation. The bar is specific. It has to state that the change is operationally safe, meaning nothing will break, no outages or service disruptions will occur, and no dependencies are affected. Something along the lines of “verified nothing uses the asset, no dependencies, safe to remove” qualifies. A thumbs up does not.
Once both verifications are in place, Tamnoon executes the fixes. The compounding effect of the previous five skills is what makes that defensible, because the verdict underneath rests on a classified asset, confirmed ownership, a documented blast radius, and verification results rather than on a model’s general confidence.
RISKY findings route to Remediation Experts who drive the complex fixes through to closure. AWAITING DATA findings pause until the missing context arrives.
Everything up to execution happens outside the live environment. Classification, prioritization, impact analysis, and the confidence verdict are all complete before anything touches production.
Remediation Built Around Your Exact Cloud
Across customer environments, this model has allowed our cloud remediation platform to achieve up to a 90% reduction in critical cloud exposure within 90 days, with zero production incidents.
The second number carries more weight for most teams. Speed is only worth having if it stays switched on, and it stays on because the skills answering each question were configured against one specific environment rather than a generic one.
See what the skills surface against your own backlog. Book a demo to see how skills-based agentic remediation can help you regain control of your backlog.
