Sweet Security

• A larger platform in the same category could add runtime coverage for AI agents to a suite it already sells, and a PeerSpot reviewer says he weighed Sweet against Wiz and Palo Alto before choosing it.
• The sixfold recurring-revenue and tenfold enterprise customer figures come from Sweet's own announcement, and no cited third party confirms either one.
• Sweet announced FedRAMP Moderate as a pursuit in June 2026 with Coalfire engaged as assessor, and no cited source records the authorization as granted, so a federal buyer that requires an authorized cloud service has to check the current status directly.
• An independent buyer guide reports that reporting and alert customization lag established platforms and that access-control permissions need refinement, which a buyer with strict access-control requirements would test before committing.
• No completed attestation report or trust portal appears in the cited sources, and a probe of trust.sweet.security on 2026-08-28 returned the marketing homepage, so a buyer will have to request assurance documents directly.

Sweet Security sells to the security leader who owns production cloud risk. SiliconANGLE describes the platform as helping teams shut down cloud attacks when and where they occur with minimal business disruption, and the site groups its solutions under CISOs, cloud security and DevSecOps, SOC and incident response, and application security. An independent buyer guide places the product with mid-market and enterprise organizations in regulated industries that prioritize catching threats in production over managing vulnerability backlogs.

The problem widened when enterprises put AI agents into production. SiliconANGLE's 2026 reporting states the case in its own voice: agents hold their own identities and reach into sensitive systems, often with no human reviewing the individual actions they take, and detection tools describe what an agent has already finished doing. Sweet sells one platform across conventional workloads and those agents.

What the cited record does not carry is a number for the pain. No source here sizes the cost of runtime cloud attacks or of agent misbehavior, so the problem stays qualitative even where two independent outlets frame it the same way. [s6, s8, s1, s10]

Sweet runs a Runtime CNAPP with a documented surface and an eBPF sensor at its base. SiliconANGLE reported in 2024 that the runtime sensor deploys in under five minutes, and an independent buyer guide that tested 10 CNAPP platforms reports the sensor needs under 100 MB of memory and builds a behavioral baseline the analytics compare against. CTech describes full visibility into systems at all stages of code execution.

The AI Security Platform extends that base to models and agents. Sweet builds a bill of materials for every model, detects agents running in the environment including shadow or unmanaged ones, and funnels agent traffic through an AI gateway that analyzes prompts and blocks malicious operations. Enforcement happens at the level of a single operation, so Sweet can revoke one risky action without ending the session, and the July 2026 release added termination of unauthorized tool calls and blocking of data leaving through an agent.

Independent validation now exists and it cuts both ways. The buyer guide credits the runtime-first approach and the incident narratives, and it reports that reporting and alert customization lag established platforms and that access-control permissions need refinement. A U.S. patent granted in December 2025 and assigned to Sweet Security Ltd covers using a language model over tokenized cloud session logs. Sweet separately credits a patented, LLM-driven detection engine with reducing alert noise to 0.04%, which is the company's own measurement. [s6, s10, s2, s12, s13, s3]

Sweet competes for the CNAPP budget against platforms its buyers also evaluate. A PeerSpot reviewer says outright that he weighed Wiz and Palo Alto before choosing Sweet, and the independent buyer guide lists Sweet alongside Prisma Cloud, Wiz, Orca Security, Sysdig and CrowdStrike in the same category. Sweet's answer is a runtime-first read of the same problem, catching what happens after an application is live rather than scanning before it ships.

The AI line moves part of that contest onto newer ground. Sweet puts a gateway in the path of agent operations and enforces least privilege as an action runs, which is a different position from watching agent traffic go by. Whether an incumbent can match that position quickly is not settled by any cited source.

Buyer research is moving toward Sweet from a small base. PeerSpot puts its CNAPP mindshare at 1.5% in August 2026, up from 0.5% a year earlier, and ranks it 14th in that category. The company is closer to the edge of the field than to its center. [s9, s10, s2]

Named customers now appear on the record with titles attached. The homepage carries Cast & Crew, ShipStation, Fireblocks, Kaltura and Engageli, quoting Fireblocks and Kaltura directly, the case-study index adds Lemonade, Firebolt and Hippo, and SiliconANGLE names Zoomd as an early customer of the agent-blocking capability. Eight named organizations appear across those two surfaces.

Two sources outside the vendor corroborate that customers run the product. PeerSpot carries six reviews rating Sweet 8.6 out of 10, all six saying they would recommend it, and one reviewer describes the runtime and application-layer depth as the place traditional platforms are weakest. An independent buyer guide reports customer feedback of its own. Neither speaks to how many customers Sweet has.

Scale remains Sweet's own claim. The Series B announcement reports a sixfold increase in annual recurring revenue, a tenfold rise in enterprise customers and multiple Fortune 1000 organizations, and the July 2026 release reports more than a billion runtime events analyzed daily. Every one of those figures is the company's own measurement, and SiliconANGLE attributes the same billion-event number to Sweet rather than reporting it independently. [s1, s15, s9, s3, s8]

The three founders carry documented senior roles in offensive and defensive cyber. SiliconANGLE describes Dror Kashti as a retired brigadier general and former chief information security officer of the Israel Defense Forces, Eyal Fisher as the former head of the cyber department at Unit 8200, and Orel Ben-Ishay as the former head of the cybersecurity research and development center at Unit 81. CTech's own account of the Series B describes the company as led by ex-IDF cyber chiefs.

The bench around them carries senior security operators. TechCrunch reports that Gerhard Eschelbeck, a former CISO at Google, and Travis McPeak, who led product security at Databricks, invested in the seed round, and the about page lists both as advisors alongside JPMorgan Chase's head of cloud security and a former director general of Israel's National Cyber Directorate. Those people advise rather than build the product.

The military pedigree is verifiable and the commercial track record is not. No cited source establishes a prior product built, a prior exit, or a sustained public research record under any founder's name. [s8, s4, s5, s7]

Sweet asks for deep access, and the assurance record in the cited pages is thin. The product holds runtime access to a customer's cloud workloads, identities and now agent operations, and the AI platform page publishes AICPA SOC 2, ISO and NIST badge images. No completed attestation report, audit letter or trust portal appears in the cited sources, and a probe of trust.sweet.security on 2026-08-28 returned the marketing homepage rather than a trust surface.

The federal move is the clearest signal of where the assurance work is going. Sweet announced in June 2026 that it is pursuing FedRAMP Moderate authorization and has engaged Coalfire Systems, a third-party assessor, to benchmark its federal security program. No cited source records that authorization as granted.

The technical on-ramp is light. SiliconANGLE reports that the sensor deploys in under five minutes, and an independent guide puts its memory footprint under 100 MB. Data handling, retention and attestation terms still have to be settled in a formal review. [s2, s14, s11, s6, s10]

Add analyzed competitors to compare them side by side with Sweet Security.

Sweet Security sells to the security organization that owns live cloud risk. SiliconANGLE describes the platform as shutting down cloud attacks when and where they occur with minimal business disruption, and the site groups its solutions under CISOs, cloud security and DevSecOps, SOC and incident response, and application security. An independent buyer guide places Sweet with regulated mid-market and enterprise teams that prioritize runtime detection and response, naming finance, healthcare and retail.

Six people speak for customers in the cited record, and four of them are security chiefs. Fireblocks, Kaltura, Cast & Crew and ShipStation are represented that way on the homepage, while Engageli speaks through a VP of research and development and SiliconANGLE quotes Zoomd's chief technology officer. PeerSpot adds that large enterprises make up 35% of the people researching Sweet on its platform, with professionals from financial services firms accounting for 9% of the views.

The AI line reaches a second audience without a second budget on the record. Sweet describes the AI platform as serving teams that build with foundation models and orchestration frameworks, and the cited pages never say who signs for the agent controls. Which function pays is unresolved.

The AI platform organizes around knowing every model, seeing every agent and controlling every action. Sweet builds a bill of materials covering every model, public or fine-tuned, across an ecosystem it describes as spanning Bedrock, Claude and OpenAI plus orchestration frameworks such as LangChain and Vertex AI. It detects agents running in the environment, including shadow or unmanaged ones, and funnels agent traffic through an AI gateway that analyzes prompts and blocks malicious operations.

Enforcement at the level of a single operation is what separates this line from watching agent traffic go by. Sweet links its workload sensors to the gateway so it can revoke one risky action in real time without ending the session, arguing that ordinary gateways see traffic rather than intent. The July 2026 release added termination of unauthorized tool calls and blocking of secrets and personal data leaving through an agent, and SiliconANGLE reported the capability in its own words.

A granted patent anchors the detection claim in a record outside Sweet's own pages. US12495057B1, assigned to Sweet Security Ltd and published in December 2025, claims a method that tokenizes cloud session logs and feeds them to a language model. Sweet separately credits a patented, LLM-driven detection engine with reducing alert noise to 0.04%, a figure that remains the company's own measurement.

What the cited record does not show is any pooling of runtime telemetry. SiliconANGLE reports Sweet's claim that its reasoning layer analyzes more than a billion runtime events daily, and the 2024 SiliconANGLE account describes sensors streaming application data to profile workload behavior. No cited page describes a cross-customer corpus, a right to reuse customer telemetry, or shared models trained on it.

Sweet sells through a demo, with a fast technical on-ramp. SiliconANGLE reports a runtime sensor that deploys in under five minutes, an independent guide puts its memory footprint under 100 MB, and every call to action on the homepage asks for a demo or a 30-minute walkthrough, with no self-serve signup on the cited pages. The technical barrier to a first look is low.

Public references are now broad. Five customers speak on the homepage with names and titles, the case-study index carries Engageli, Kaltura, Fireblocks, Firebolt, Lemonade and Hippo, and SiliconANGLE names Zoomd as an early customer of the agent-blocking release. PeerSpot carries six reviews rating Sweet 8.6 out of 10 with all six saying they would recommend it, which corroborates use from outside the vendor.

Scale itself is still Sweet's own number. The Series B announcement reports a sixfold increase in annual recurring revenue and a tenfold rise in enterprise customers including multiple Fortune 1000 organizations, and CTech reports that around $15 million of the round went to secondary deals rather than into the company. SiliconANGLE relays the billion-events figure as something the company said, so no cited source confirms any growth or volume claim independently.

Sweet publishes no price and no billing unit in the cited pages. The homepage asks for a demo and a 30-minute walkthrough, the integrations catalog and platform pages carry no packaging tiers, and an independent buyer guide that publishes a starting price for two rivals records Sweet as contact for quote with billing not disclosed. A buyer cannot anchor a budget before talking to sales.

The platform shape suggests value scales with the estate under protection. Sweet covers detection and response, posture, vulnerability management, identity, API security, data security and the AI line from one product, and an outside reader cannot tell whether it charges by workload, sensor, asset or tier. The cited record stops short of the meter.

A buyer therefore weighs Sweet on references and a demo rather than on a comparable quote. That fits the enterprise segment the independent guide describes, and it puts more weight on the named customers and the six PeerSpot reviews than a published price list would.

Sweet ships software the customer deploys and operates. The eBPF sensor installs in under five minutes, needs under 100 MB of memory, and streams application data to profile workload behavior, and Sweet extended that sensor to Windows environments in October 2025. Nothing in the cited record describes Sweet operating the product on a customer's behalf.

The AI line adds an enforcement component that stands in the path of agent actions. Sweet links workload sensors to the gateway for contextual, operation-level enforcement and blocks unauthorized actions without ending the session. SiliconANGLE names the operational hazard plainly, that enforcement misreading legitimate behavior takes down production, and reports Sweet's argument that this risk is why most of the industry has settled for alerting instead.

Sweet publishes no operating commitments. No uptime figure, support commitment or failure-mode description appears in the cited pages, which is what a careful review asks for before putting an enforcement layer in front of production agents.

Sweet asks for deep access, and the assurance record in the cited pages is thin. The AI platform page carries badge images labelled AICPA SOC 2, ISO and NIST, and no completed attestation report, audit letter or trust portal appears in the cited pages. A probe of trust.sweet.security on 2026-08-28 returned the marketing homepage rather than a trust surface.

The federal move is where the assurance work is going. Sweet announced in June 2026 that it is pursuing FedRAMP Moderate authorization and has engaged Coalfire Systems, a third-party assessor, to benchmark its federal security program, and its CPO frames the work as continued investment in operational rigor and governance. No cited source records that authorization as granted.

Sweet makes the rest of its trust case on architecture and people. It positions runtime context plus operation-level enforcement as the route to precise real-time protection, and its founders and advisors carry documented security careers. A buyer routing privileged cloud and agent activity through Sweet still has to settle data handling, retention and attestation terms in a formal review.

Sweet positions itself as one platform across conventional cloud and AI. The Runtime CNAPP page lists detection and response, identity, application and API security, vulnerability management, posture, data security, entitlements and compliance checks under one product, and CTech describes a comprehensive CNAPP spanning the stages of code execution. Sweet frames that breadth as consolidation, calling itself ADR, CDR and CWPP combined.

Outward reach comes through two channels the cited pages document. The integrations catalog lists categories for CI/CD, cloud providers, code security, notifications, privileged access, security workflows, SIEM, SSO, third-party data and ticketing, with named entries such as CyberArk. The AI line reaches outward instead through the gateway that sits where agents call tools, which Sweet argues is where intent forms.

The exposure is bundling by a larger platform in the same category. A PeerSpot reviewer says he evaluated Wiz and Palo Alto before choosing Sweet, and an independent guide lists Sweet alongside Prisma Cloud, Wiz, Orca Security, Sysdig and CrowdStrike in the same category. The cloud coverage is the half those platforms contest most directly.

Three founders with documented senior cyber careers run the company. SiliconANGLE describes Dror Kashti as a retired brigadier general and former chief information security officer of the Israel Defense Forces, Eyal Fisher as the former head of the cyber department at Unit 8200, and Orel Ben-Ishay as the former head of the cybersecurity research and development center at Unit 81. CTech independently describes the company as led by ex-IDF cyber chiefs. Tomer Filiba is named as CTO and the cited pages state no prior background for him.

The advisory bench carries senior security operators. SiliconANGLE reported that Gerhard Eschelbeck, a former CISO at Google, and Travis McPeak, who led product security at Databricks, were investors in the seed round, and the about page lists both as advisors alongside JPMorgan Chase's head of cloud security and a former director general of Israel's National Cyber Directorate. Those people advise the company rather than build or buy the product.

What no cited source establishes is a company any founder built before this one. The military records are specific and verifiable, and a prior product, an exit, or a sustained public research record under a founder's name does not appear in the reviewed sources.