The cloud security landscape has reached a turning point where the ability to spot misconfigurations is no longer the bottleneck; the real challenge lies in acting on those discoveries without creating new risks. For years, vendors poured energy into sharpening scanners, ensuring that an exposed bucket, an open port, or an overโprivileged role surfaces in a console within minutes of appearing. This nearโrealโtime visibility has become table stakes, and few dispute that detection work is essentially complete. Yet the industryโs own metrics reveal a stark contradiction: more than half of every finding ever logged remains unresolved, and the average time to close a critical issue now stretches to about five monthsโa window that dwarfs the minutes attackers need to move from initial foothold to lateral spread. The gap between seeing a problem and fixing it has become a liability, turning security teams into archivists of risk rather than removers of threat. Leaders who once celebrated the completeness of their telemetry now stare at mounting backlogs and wonder why the promised speed of automated fixes has not materialized. The answer, as emerging evidence shows, is not merely a matter of faster buttons or smoother approval workflows; it rests on a deeper question of trustโwhether a machine can be relied upon to intervene only when it is truly safe to do so.
Marina Segal brings a unique vantage point to this debate, having helped construct the very scanners that now flood security consoles with alerts. Early in her career she contributed to Dome9 Security, where the concept of cloud security posture management was first codified and later acquired by Check Point, solidifying the industryโs focus on continuous visibility. After the acquisition she remained inside the enlarged organization, watching as the flood of alerts grew while the corresponding remediation effort stagnated. This firsthand experience gave her a clear view of the disconnect: teams could pinpoint every misconfiguration, yet the tickets generated by those findings lingered untouched, filed away like paperwork in a cabinet. Rather than accepting the status quo as an inevitable byproduct of scale, Segal chose to interrogate the underlying assumption that more automation equals better outcomes. She hypothesized that the industryโs relentless push for speed overlooked a fundamental human concernโaccountability. In production environments, every change bears a name, and stakeholders are loath to endorse an alteration they cannot justify if something goes wrong. This insight set the stage for a new kind of security tool, one that would prioritize restraint over rash action and seek to earn trust through demonstrable caution rather than raw speed.
The numbers Segalโs own team has gathered paint a quantitative picture of the remediation deficit. Across the tens of millions of detections processed by their platform, more than 50โฏpercent of findings stay open indefinitely, turning the cloud into a repository of known weaknesses. A critical misconfiguration, on average, requires roughly 150 days to healโa period that has been creeping upward rather than downward. To put that interval in perspective, threat intelligence from CrowdStrike indicates that adversaries can progress from an initial compromise to lateral movement in under half an hour. Consequently, once an alert fires, the defenderโs clock runs in months while the attackerโs operates in minutes, creating a dangerous asymmetry that favors the aggressor. This mismatch is not a failure of detection fidelity; scanners consistently flag the same issues with high confidence. Instead, the bottleneck emerges after the alert, where human teams wrestle with uncertainty about ownership, dependencies, and potential side effects. The resulting hesitation is rational: a mistaken remediation can masquerade as an operational outage, complete with a ticket that points directly to the person who approved the change. Until a solution can consistently demonstrate that its interventions are safe, the natural reluctance to automate will persist, leaving valuable security signals stuck in a perpetual queue.
The marketโs response to the remediation backlog has largely taken the shape of increasingly aggressive automation promises. Virtually every cloud security platform now ships an agent that claims to close findings autonomously, highlighting features such as rapid execution, an optional approval step, rollback safeguards, and granular control panels. Marketing materials trumpet speed as the primary benefit, suggesting that the only missing ingredient is a faster button to push. In practice, however, the majority of these autonomous switches remain deactivated in production environments. Security leaders admit that the technology itself works; the hesitation stems from a deeper cultural and procedural concern: when a machine proposes a change, the ensuing ticket bears the name of the human who signed off, and any ensuing disruption is attributed to that individual. Consequently, teams prefer to keep the automation feature turned off, opting instead for manual remediation despite its inefficiency, because the manual route preserves a clear chain of accountability. This dynamic reveals that the industry has been solving the wrong problemโaccelerating the act of remediation while neglecting the prerequisite of confidence that the action will not cause collateral damage. Without a mechanism to prove safety, speed alone becomes a liability rather than an advantage.
The apprehension surrounding automated fixes is not a matter of insufficient training, inadequate changeโmanagement processes, or resistance to new tools; it is a perfectly rational reaction to systems that have never demonstrated an ability to refuse a risky proposal. When a vendor presents an automation feature that can only say โyesโ, the underlying message is that the tool will attempt to remediate every finding it encounters, regardless of context. In a production setting, that blind optimism translates into potential service interruptions, data loss, or compliance violations, all of which would be traced back to the human who authorized the machineโs action. Segal articulates this concern succinctly: a machine incapable of saying โnoโ is not an automation asset but a liability wrapped in an API. The fear is therefore justified; stakeholders are not afraid that the fix will be too slow, but that it might be wrong, and that they will be held responsible for endorsing an erroneous change. Until automation vendors can exhibit a track record of prudent refusalsโcomplete with the reasoning behind each negative decisionโthe market will remain skeptical, and the promised efficiency gains will stay locked behind manual processes.
Tamnoonโs engineering philosophy inverts the conventional automation narrative by starting with restraint rather than action. Before the engine is permitted to alter any resource in a customerโs environment, it undergoes a rigorous learning phase focused on understanding context: what workloads depend on the target asset, who owns it, what recent changes have occurred, and what network traffic patterns suggest about its role. Only after this comprehension phase does the system begin to evaluate whether a proposed remediation is safe, and it is trained to refuse the change whenever sufficient evidence of safety is lacking. This sequencing means that the first automated competency the model acquires is the ability to decline, not to apply a fix. Segal emphasizes that this orderโteaching when to stop before teaching how to actโis foundational to the companyโs value proposition. By prioritizing cautious judgment, Tamnoon aims to build trust incrementally, proving that its interventions are grounded in a deep appreciation of the operational fabric rather than a blunt willingness to alter configuration settings.
To illustrate the nuance of its decisionโmaking, Tamnoon runs a brief demonstration that takes less than a minute to observe. Six identicalโlooking Amazon S3 buckets are presented to the engine, each flagged by scanners as open to the internet and assigned the same severity level. The scanner, unable to differentiate beyond the surface configuration, treats all six as equally risky. Tamnoonโs model, however, examines each bucket in light of additional signals. Three of the buckets are determined to be safe to close; the engine proceeds to shut them down through the customerโs native changeโmanagement workflow, complete with a preโconfigured rollback plan. Two buckets are declined outright because closing them would disrupt a live dependency that the scanner cannot seeโperhaps a legacy application pulling data from the bucket or a logging stream that relies on public read access. The remaining bucket triggers a request for clarification: the ownership tag is missing or ambiguous, and the system refuses to guess, instead inviting a human to supply the missing context. Thus, from a single alert type the engine produces three distinct, correct outcomesโtwo closures, two refusals, and one inquiryโshowing that the scannerโs accuracy was never the issue; the missing layer was contextual judgment.
A refusal that lacks justification is little more than a shrug and does little to alleviate the anxiety of security teams. Tamnoon ensures that every negative decision is accompanied by a detailed evidence package: the original finding, the specific data points the engine consulted, the verdict, a concise rationale, the single fact that would flip the decision to a โyesโ, and stepโbyโstep instructions for manual remediation if the customer chooses to act. This transparency transforms a simple โnoโ into a defensible position that can be presented at a morning standโup, audited by compliance officers, or reviewed by executive leadership. By providing the evidence that would change its mind, the engine invites collaboration rather than confrontation; stakeholders can see exactly what additional informationโsuch as a verified owner tag or a confirmed lowโtraffic windowโwould permit the change to proceed safely. Consequently, the refusal becomes a constructive signal, guiding teams toward the precise remediation path that balances security with operational continuity.
Where many vendors treat autonomy as a configurable dial that customers can turn up or down, Tamnoon adopts an earningโbased model that begins with zero privileges. Upon initial connection, the engine operates in a strictly readโonly mode, gathering observations without attempting any modifications. Only after it demonstrates a consistent record of safe and correct decisions on a particular class of findingโsuch as open storage bucketsโis it granted limited authority to act on that class, and even then the customer retains an overriding veto that can be exercised at any time. This approach mirrors the way financial institutions extend credit: a limit is increased only after a proven history of responsible usage. Segal likens autonomy to a track record that can be inspected, not a feature purchased off the shelf. Should the engine ever make an erroneous call, the oversight mechanism allows the customer to revoke its permissions, reinforcing that trust is continuously negotiated rather than assumed. Over time, the systemโs permissions evolve in lockstep with demonstrated reliability, ensuring that any escalation in capability is justified by empirical evidence rather than marketing promises.
Every decisionโwhether a closure, a refusal, or a request for clarificationโis immortalized in a tamperโevident receipt that consists of six standardized fields: the finding identifier, the contextual data evaluated, the verdict, the justification, the mutable condition that would alter the outcome, and the prescribed manual steps. These receipts are designed to be readable by auditors without requiring specialized translation, thereby facilitating compliance reviews and internal investigations. When a human operator overrides the engineโs suggestion, that override is recorded as a new receipt and fed back into the modelโs learning pipeline, allowing the system to refine its judgments for that specific environment. Consequently, the engine becomes smarter not only from aggregate fleet data but also from the unique nuances of each customerโs cloud, growing stricter where needed and more permissive where proven safe. This closedโloop feedback transforms the security stack from a static set of rules into an evolving knowledge base that adapts to both shifting threat landscapes and evolving operational practices.
Segalโs outlook for the next halfโdecade paints a picture in which security teams transition from being the gatekeepers of every configuration change to becoming curators of vetted, lowโrisk remediations. The traditional backlog, once a static filing cabinet of unresolved findings, would morph into a dynamic queue where each entry carries a clear verdict and supporting evidence, enabling rapid triage. Engineers would no longer face hundreds of redundant alerts for the same underlying issue; Tamnoonโs crossโscanner correlation model shows that the average organization sees roughly one actionable problem for every 142 raw detections, dramatically reducing noise. Because the engine only approves changes that it can prove safe, healed configurations tend to stay healed, preventing the recurrence of the same eightโtoโten misconfigurations that plague quarterly reports. Autonomy, rather than being a blanket toggle in a vendorโs settings panel, would evolve into a nuanced privilegeโcomparable to a credit limitโthat is allocated per finding class only after the engineโs track record satisfies predefined safety thresholds. In this model, the boardโs line of questioning shifts from โHow fast was the fix?โ to โHow did you know it was safe?โ, a question that Tamnoonโs receipts are expressly built to answer.
Three concrete predictions accompany this vision, and Segal invites the industry to hold her accountable should they fail to materialize. First, the ability to demonstrate a refusalโcomplete with evidenceโwill become a mandatory line item in every remediationโfocused request for proposal, shifting evaluation criteria from speed alone to prudence. Second, the receipt format will supplant the traditional score or metric as the primary artifact buyers compare when weighing competing solutions, because it directly communicates decision quality. Third, no reputable vendor will be able to sell a remediation product without showcasing at least one example of a change it declined, complete with the reasoning that led to that outcome. For security leaders preparing to evaluate or renew their cloudโsecurity stack, the practical advice is straightforward: insist on seeing a documented refusal before committing to any automation feature, evaluate whether the vendor supplies transparent evidence for each negative decision, and prioritize solutions that tie expanding autonomy to a verifiable track record rather than to a simple configuration switch. By adopting this evidenceโfirst mindset, organizations can close the remediation gap, reduce meanโtimeโtoโheal from months to hours, and transform their cloud security posture from a reactive archive into a proactive, trustworthy safeguard.