Sweet Security Unveils AI Red-Teaming Agent to Predict Cloud Attack Paths

The cloud security landscape just got a lot more proactive. Sweet Security, a firm known for its focus on runtime protection, has officially pulled the curtain back on a new AI-powered red-teaming agent. This tool isn't just another scanner; it is designed to think like an adversary, mapping out potential "attack paths" that a human hacker might take to compromise a corporate network.

Traditional security measures often rely on static rules or historical data to flag threats. However, as cloud environments become more complex, the number of ways a system can be breached grows exponentially. Sweet’s new agent utilizes generative AI to simulate these breaches in real-time, providing security teams with a much-needed "attacker’s eye view" of their infrastructure.

According to reports from Dark Reading, the agent works by analyzing the specific configuration and behavior of a company's cloud stack. By understanding how different services interact, the AI can identify non-obvious links that could be chained together to escalate privileges or exfiltrate sensitive data.

Moving Beyond Static Vulnerability Management

The core problem Sweet is addressing is "alert fatigue." Security practitioners are often buried under thousands of low-priority vulnerabilities. By using an AI red-teamer, the platform can prioritize risks based on exploitability. If a vulnerability exists but is unreachable via any realistic attack path, the AI helps the team deprioritize it in favor of more critical, "path-connected" threats.

This shift represents a move toward continuous security validation. Rather than waiting for a biannual manual penetration test, companies can essentially run a permanent, automated "war game" within their environment. This ensures that as new code is deployed or configurations are changed, the security posture is updated instantly.

The timing of this launch is significant. As highlighted by SC Media, the rise of "living off the land" attacks—where hackers use legitimate system tools to conduct malicious activities—has made traditional detection much harder. An AI that understands behavioral patterns is better equipped to spot these subtle anomalies.

The Architecture of Autonomy

Technically, the agent operates within the Sweet Security Runtime Suite. It leverages large language models (LLMs) specialized in cybersecurity logic to predict the next steps an attacker might take. This allows the system to not only report a flaw but also demonstrate the potential business impact by showing exactly which assets are at risk.

For CISOs, the value proposition is clear: efficiency. Manual red-teaming is expensive and time-consuming. While an AI won't replace a highly skilled human specialist entirely, it can handle the "brute force" work of checking thousands of permutations, leaving humans to focus on the most complex strategic defenses.

Industry analysts have noted that this move aligns with a broader trend in the industry. As noted by VentureBeat, the integration of AI into the defensive side of cybersecurity is a necessary response to the fact that attackers are already using AI to automate their own reconnaissance and malware development.

Closing the "Detection Gap"

One of the biggest hurdles in cloud security is the "gap" between seeing a vulnerability and understanding if it actually matters. Sweet’s agent aims to close this gap by providing proof of concept. If the AI can reach a database from a public-facing web server, the security team knows they have a verified fire to put out.

Furthermore, the agent's ability to operate in "runtime" is a major differentiator. Many cloud security tools focus on the "shift left" philosophy—securing code before it’s deployed. Sweet argues that while pre-deployment security is vital, the real danger lies in how applications behave once they are live and interacting with the real world.

The feedback loop created by this AI agent also helps in refining automated response playbooks. When the red-teaming agent discovers a new path, the system can automatically suggest—or in some cases, implement—micro-segmentation rules to cut that path off before a real-world threat actor finds it.

Looking Toward an AI-First Defense

As we move further into 2024 and beyond, the "arms race" between AI-powered attackers and AI-powered defenders will only intensify. Sweet Security’s launch is a stake in the ground, suggesting that the future of the SOC (Security Operations Center) is one where autonomous agents handle the heavy lifting of threat modeling.

Ultimately, the success of such tools will depend on their accuracy and their ability to avoid "hallucinations"—a common pitfall for AI. If the red-teamer flags too many "phantom" attack paths that aren't actually possible, it risks contributing to the very noise it was designed to silence.

However, for now, the industry seems optimistic. By turning the "black box" of cloud infrastructure into a transparent map of risks, Sweet Security is giving defenders a fighting chance to stay one step ahead of the curve. It’s a bold move that signals a new era of proactive, intelligent cloud defense.

The Backstory: A Mission Forged in High-Stakes Cloud Defense

The emergence of the "Sweet Attack" agent is not a sudden pivot but the culmination of a journey that began within the highest levels of the Israeli military. Founded in 2023, Sweet Security was established by a trio of veterans from Israel’s elite technological intelligence units. Leading the charge is CEO Dror Kashti, a retired Brigadier General who previously served as the Chief Information Security Officer (CISO) for the Israel Defense Forces (IDF). His experience managing large-scale cloud migrations—specifically the high-profile Project Nimbus—exposed the critical limitations of existing security tools when faced with the dynamic nature of runtime environments.

Kashti is joined by Co-Founder and CPO Eyal Fisher, a retired Colonel who led the Cyber Department at the legendary Unit 8200, often referred to as Israel's equivalent to the NSA. Rounding out the founding team is VP of R&D Orel Ben-Ishay, who previously headed the cybersecurity research and development center at Unit 81. This pedigree has allowed Sweet Security to infuse its platform with an "adversarial mindset," treating cloud defense not as a static checklist, but as a live battlefield where understanding attacker behavior is paramount.

The company’s growth trajectory has been equally rapid. Just six months after emerging from stealth with a $12 million seed round, Sweet Security secured an additional $33 million in Series A funding in early 2024. More recently, in November 2025, the firm announced a massive $75 million Series B round led by Evolution Equity Partners. This latest influx of capital, which brought their total funding to $120 million, was explicitly earmarked to accelerate global expansion and the development of the autonomous AI capabilities seen in their new red-teaming agent.

Technology Grounded in "Ground Truth"

What sets the Sweet Attack agent apart technically is its reliance on what the company calls "ground truth" data. Most security tools operate by taking snapshots of cloud configurations or scanning static code. In contrast, Sweet’s platform utilizes a lightweight sensor based on eBPF (Extended Berkeley Packet Filter) technology. This allows the system to monitor the actual, live execution of code, network traffic at Layer 7, and the behavior of non-human identities without imposing significant overhead on the system’s performance.

This deep visibility provides the substrate upon which the AI agent reasons. While a standard LLM might only be able to hypothesize about general vulnerabilities, the Sweet Attack agent uses the live index of the customer's specific environment to find real, exploitable paths. It essentially combines the broad reasoning of a frontier model with the specific, technical context of a company's production environment, ensuring that the simulated attacks are both realistic and reproducible.

The urgency of this technology is highlighted by the "Mythos" benchmark, a new standard cited by SecurityBrief. As security teams face increasing pressure to quantify their exposure to AI-assisted threats, the Sweet Attack agent provides a measurable way to test defenses. By mapping out every possible chain—from an exposed API to a sensitive database—it helps organizations move from a reactive "catch the hacker" mode to a proactive "close the path" strategy.

Impact on the Modern Security Operations Center

The broader goal for the Sweet team is to achieve a "99% reduction in critical vulnerability noise." In today’s cloud environments, developers are often overwhelmed by "CVE (Common Vulnerabilities and Exposures) spam." Many of these vulnerabilities are theoretically risky but practically unreachable. By using an autonomous red-teaming agent to verify exploitability, Sweet allows teams to ignore the noise and focus on the handful of risks that a determined attacker could actually reach.

This approach also addresses the "skills gap" in the cybersecurity industry. Sophisticated penetration testing is a rare and expensive skill. By automating the adversarial probing process, SecurityWeek notes that Sweet is democratizing high-level defense strategies, allowing even smaller security teams to benefit from the same level of rigorous testing previously reserved for the most well-funded enterprises.

Looking ahead, Sweet Security aims to be the dominant leader in what is now being called "Runtime CNAPP" (Cloud-Native Application Protection Platform). As cloud environments and the AI systems running on them become increasingly autonomous, the tools used to protect them must follow suit. The launch of this red-teaming agent is more than just a product update; it’s a vision of a future where security is a self-testing, self-healing organism that evolves as quickly as the threats it faces.

Reading Between the Lines: The Shift from Defensive Accounting to Proactive Warfare

The launch of Sweet’s AI red-teaming agent marks a philosophical pivot in the cybersecurity industry: we are moving away from "vulnerability accounting" toward "exploitability validation." For years, the industry has focused on the volume of flaws—counting CVEs like beans in a jar. But in a cloud-native world where resources are ephemeral and connections are fluid, the sheer number of flaws is a poor metric for actual risk. By deploying an autonomous agent that actively seeks out attack paths, Sweet is effectively arguing that a vulnerability doesn't truly exist unless an attacker can actually reach it.

This "adversarial-first" logic reflects a growing realization that human-led security cannot scale with the speed of automated infrastructure. As companies adopt Infrastructure-as-Code (IaC) and CI/CD pipelines, the cloud environment changes minute-by-minute. A manual penetration test performed on a Monday might be obsolete by Tuesday afternoon. Sweet’s agent introduces a layer of "continuous friction" for attackers, forcing them to contend with a defensive system that is constantly probing its own perimeter and closing doors before they can be kicked in.

From a market perspective, this move puts immense pressure on traditional Cloud Security Posture Management (CSPM) vendors. Many legacy tools provide visibility but lack the "teeth" of runtime intervention. By integrating red-teaming into a runtime suite, Sweet is bridging the gap between identifying a problem and proving its lethality. This convergence of offensive and defensive AI suggests that the next generation of security platforms will not just be dashboards, but active participants in the digital battlefield.

The "Hallucination" Hurdle in Security Logic

However, the transition to agentic AI in security is not without its risks. The most significant technical challenge lies in the precision of the AI’s reasoning. In a security context, an AI "hallucination"—where the agent identifies a false attack path or, worse, misses a real one—can lead to a false sense of security. Sweet's reliance on eBPF data is a strategic move to mitigate this, providing the AI with hard, behavioral evidence rather than just configuration files to "guess" from.

Moreover, there is the question of "AI vs. AI" escalation. As defenders use agents like Sweet’s to harden their stacks, attackers will inevitably use similar agents to find "shadow paths" that bypass known logic. This creates a recursive loop where the complexity of the security environment increases alongside the sophistication of the tools. The industry must be careful not to create a system so complex that human operators can no longer audit the AI’s decisions, leading to a "black box" security posture.

There is also a significant cultural shift required within IT teams. Historically, "red-teaming" was a disruptive event that developers often viewed with suspicion. Integrating an autonomous red-teamer into production runtime requires a high level of trust. If the agent is too aggressive, it risks impacting performance; if it is too passive, it becomes shelfware. Sweet's challenge will be maintaining that delicate balance between rigorous testing and operational stability.

Economics of the Autonomous SOC

The economic implications are equally profound. The current "CISO's dilemma" is that security budgets are finite while the attack surface is infinite. By automating the most labor-intensive parts of the security lifecycle—reconnaissance and path analysis—Sweet is offering a path toward a more sustainable ROI. Instead of hiring ten more analysts to sift through logs, a company might hire two highly skilled "AI orchestrators" to manage a fleet of autonomous agents.

This shift could lead to a consolidation of the security stack. If one platform can handle monitoring, red-teaming, and incident response, the need for dozens of disparate "point solutions" diminishes. We are likely seeing the early stages of a "platform war" where the winners will be those who can provide a unified, autonomous loop: see the threat, simulate the path, and block the execution, all in sub-seconds.

Finally, we must consider the regulatory landscape. As governments around the world begin to mandate more rigorous cybersecurity testing, tools that provide "automated proof of defense" will become essential. Sweet is positioning itself not just as a security provider, but as an automated auditor. In the future, a company might prove its compliance not with a static report, but by showing that its autonomous red-teaming agent failed to find a single path to its crown jewels over the previous quarter.

"In the end, we’re essentially teaching our cloud environments to be their own worst critics—which is great for security, but hopefully, the AI doesn't get too good at it, or we'll all be locked out of our own servers by a robot that thinks we're the 'weakest link' in the attack path."

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn