Blog
/
Network
/
September 5, 2022

Steps of a BumbleBee Intrusion to a Cobalt Strike

Discover the steps of a Bumblebee intrusion, from initial detection to Cobalt Strike deployment. Learn how Darktrace defends against evolving threats with AI.
Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Sam Lister
Specialist Security Researcher
Default blog image
05
Sep 2022

Introduction

Throughout April 2022, Darktrace observed several cases in which threat actors used the loader known as ‘BumbleBee’ to install Cobalt Strike Beacon onto victim systems. The threat actors then leveraged Cobalt Strike Beacon to conduct network reconnaissance, obtain account password data, and write malicious payloads across the network. In this article, we will provide details of the actions threat actors took during their intrusions, as well as details of the network-based behaviours which served as evidence of the actors’ activities.  

BumbleBee 

In March 2022, Google’s Threat Analysis Group (TAG) provided details of the activities of an Initial Access Broker (IAB) group dubbed ‘Exotic Lily’ [1]. Before March 2022, Google’s TAG observed Exotic Lily leveraging sophisticated impersonation techniques to trick employees of targeted organisations into downloading ISO disc image files from legitimate file storage services such as WeTransfer. These ISO files contained a Windows shortcut LNK file and a BazarLoader Dynamic Link Library (i.e, DLL). BazarLoader is a member of the Bazar family — a family of malware (including both BazarLoader and BazarBackdoor) with strong ties to the Trickbot malware, the Anchor malware family, and Conti ransomware. BazarLoader, which is typically distributed via email campaigns or via fraudulent call campaigns, has been known to drop Cobalt Strike as a precursor to Conti ransomware deployment [2]. 

In March 2022, Google’s TAG observed Exotic Lily leveraging file storage services to distribute an ISO file containing a DLL which, when executed, caused the victim machine to make HTTP requests with the user-agent string ‘bumblebee’. Google’s TAG consequently called this DLL payload ‘BumbleBee’. Since Google’s discovery of BumbleBee back in March, several threat research teams have reported BumbleBee samples dropping Cobalt Strike [1]/[3]/[4]/[5]. It has also been reported by Proofpoint [3] that other threat actors such as TA578 and TA579 transitioned to BumbleBee in March 2022.  

Interestingly, BazarLoader’s replacement with BumbleBee seems to coincide with the leaking of the Conti ransomware gang’s Jabber chat logs at the end of February 2022. On February 25th, 2022, the Conti gang published a blog post announcing their full support for the Russian state’s invasion of Ukraine [6]. 

Figure 1: The Conti gang's public declaration of their support for Russia's invasion of Ukraine

Within days of sharing their support for Russia, logs from a server hosting the group’s Jabber communications began to be leaked on Twitter by @ContiLeaks [7]. The leaked logs included records of conversations among nearly 500 threat actors between Jan 2020 and March 2022 [8]. The Jabber logs were supposedly stolen and leaked by a Ukrainian security researcher [3]/[6].

Affiliates of the Conti ransomware group were known to use BazarLoader to deliver Conti ransomware [9]. BumbleBee has now also been linked to the Conti ransomware group by several threat research teams [1]/[10]/[11]. The fact that threat actors’ transition from BazarLoader to BumbleBee coincides with the leak of Conti’s Jabber chat logs may indicate that the transition occurred as a result of the leaks [3]. Since the transition, BumbleBee has become a significant tool in the cyber-crime ecosystem, with links to several ransomware operations such as Conti, Quantum, and Mountlocker [11]. The rising use of BumbleBee by threat actors, and particularly ransomware actors, makes the early detection of BumbleBee key to identifying the preparatory stages of ransomware attacks.  

Intrusion Kill Chain 

In April 2022, Darktrace observed the following pattern of threat actor activity within the networks of several Darktrace clients: 

1.     Threat actor socially engineers user via email into running a BumbleBee payload on their device

2.     BumbleBee establishes HTTPS communication with a BumbleBee C2 server

3.     Threat actor instructs BumbleBee to download and execute Cobalt Strike Beacon

4.     Cobalt Strike Beacon establishes HTTPS communication with a Cobalt Strike C2 server

5.     Threat actor instructs Cobalt Strike Beacon to scan for open ports and to enumerate network shares

6.     Threat actor instructs Cobalt Strike Beacon to use the DCSync technique to obtain password account data from an internal domain controller

7.     Threat actor instructs Cobalt Strike Beacon to distribute malicious payloads to other internal systems 

With limited visibility over affected clients’ email environments, Darktrace was unable to determine how the threat actors interacted with users to initiate the BumbleBee infection. However, based on open-source reporting on BumbleBee [3]/[4]/[10]/[11]/[12]/[13]/[14]/[15]/[16]/[17], it is likely that the actors tricked target users into running BumbleBee by sending them emails containing either a malicious zipped ISO file or a link to a file storage service hosting the malicious zipped ISO file. These ISO files typically contain a LNK file and a BumbleBee DLL payload. The properties of these LNK files are set in such a way that opening them causes the corresponding DLL payload to run. 

In several cases observed by Darktrace, devices contacted a file storage service such as Microsoft OneDrive or Google Cloud Storage immediately before they displayed signs of BumbleBee infection. In these cases, it is likely that BumbleBee was executed on the users’ devices as a result of the users interacting with an ISO file which they were tricked into downloading from a file storage service. 

Figure 2: The above figure, taken from the event log for an infected device, shows that the device contacted a OneDrive endpoint immediately before making HTTPS connections to the BumbleBee C2 server, 45.140.146[.]244
Figure 3: The above figure, taken from the event log for an infected device, shows that the device contacted a Google Cloud Storage endpoint and then the malicious endpoint ‘marebust[.]com’ before making HTTPS connections to the  BumbleBee C2 servers, 108.62.118[.]61 and 23.227.198[.]217

After users ran a BumbleBee payload, their devices immediately initiated communications with BumbleBee C2 servers. The BumbleBee samples used HTTPS for their C2 communication, and all presented a common JA3 client fingerprint, ‘0c9457ab6f0d6a14fc8a3d1d149547fb’. All analysed samples excluded domain names in their ‘client hello’ messages to the C2 servers, which is unusual for legitimate HTTPS communication. External SSL connections which do not specify a destination domain name and whose JA3 client fingerprint is ‘0c9457ab6f0d6a14fc8a3d1d149547fb’ are potential indicators of BumbleBee infection. 

Figure 4:The above figure, taken from Darktrace's Advanced Search interface, depicts an infected device's spike in HTTPS connections with the JA3 client fingerprint ‘0c9457ab6f0d6a14fc8a3d1d149547fb’

Once the threat actors had established HTTPS communication with the BumbleBee-infected systems, they instructed BumbleBee to download and execute Cobalt Strike Beacon. This behaviour resulted in the infected systems making HTTPS connections to Cobalt Strike C2 servers. The Cobalt Strike Beacon samples all had the same JA3 client fingerprint ‘a0e9f5d64349fb13191bc781f81f42e1’ — a fingerprint associated with previously seen Cobalt Strike samples [18]. The domain names ‘fuvataren[.]com’ and ‘cuhirito[.]com’ were observed in the samples’ HTTPS communications. 

Figure 5:The above figure, taken from Darktrace's Advanced Search interface, depicts the Cobalt Strike C2 communications which immediately followed a device's BumbleBee C2 activity

Cobalt Strike Beacon payloads call home to C2 servers for instructions. In the cases observed, threat actors first instructed the Beacon payloads to perform reconnaissance tasks, such as SMB port scanning and SMB enumeration. It is likely that the threat actors performed these steps to inform the next stages of their operations.  The SMB enumeration activity was evidenced by the infected devices making NetrShareEnum and NetrShareGetInfo requests to the srvsvc RPC interface on internal systems.

Figure 6: The above figure, taken from Darktrace’s Advanced Search interface, depicts a spike in srvsvc requests coinciding with the infected device's Cobalt Strike C2 activity

After providing Cobalt Strike Beacon with reconnaissance tasks, the threat actors set out to obtain account password data in preparation for the lateral movement phase of their operation. To obtain account password data, the actors instructed Cobalt Strike Beacon to use the DCSync technique to replicate account password data from an internal domain controller. This activity was evidenced by the infected devices making DRSGetNCChanges requests to the drsuapi RPC interface on internal domain controllers. 

Figure 7: The above figure, taken from Darktrace’s Advanced Search interface, depicts a spike in DRSGetNCChanges requests coinciding with the infected device’s Cobalt Strike C2 activity

After leveraging the DCSync technique, the threat actors sought to broaden their presence within the targeted networks.  To achieve this, they instructed Cobalt Strike Beacon to get several specially selected internal systems to run a suspiciously named DLL (‘f.dll’). Cobalt Strike first established SMB sessions with target systems using compromised account credentials. During these sessions, Cobalt Strike uploaded the malicious DLL to a hidden network share. To execute the DLL, Cobalt Strike abused the Windows Service Control Manager (SCM) to remotely control and manipulate running services on the targeted internal hosts. Cobalt Strike first opened a binding handle to the svcctl interface on the targeted destination systems. It then went on to make an OpenSCManagerW request, a CreateServiceA request, and a StartServiceA request to the svcctl interface on the targeted hosts: 

·      Bind request – opens a binding handle to the relevant RPC interface (in this case, the svcctl interface) on the destination device

·      OpenSCManagerW request – establishes a connection to the Service Control Manager (SCM) on the destination device and opens a specified SCM database

·      CreateServiceA request – creates a service object and adds it to the specified SCM database 

·      StartServiceA request – starts a specified service

Figure 8: The above figure, taken from Darktrace’s Advanced Search interface, outlines an infected system’s lateral movement activities. After writing a file named ‘f.dll’ to the C$ share on an internal server, the infected device made several RPC requests to the svcctl interface on the targeted server

It is likely that the DLL file which the threat actors distributed was a Cobalt Strike payload. In one case, however, the threat actor was also seen distributing and executing a payload named ‘procdump64.exe’. This may suggest that the threat actor was seeking to use ProcDump to obtain authentication material stored in the process memory of the Local Security Authority Subsystem Service (LSASS). Given that ProcDump is a legitimate Windows Sysinternals tool primarily used for diagnostics and troubleshooting, it is likely that threat actors leveraged it in order to evade detection. 

In all the cases which Darktrace observed, threat actors’ attempts to conduct follow-up activities after moving laterally were thwarted with the help of Darktrace’s SOC team. It is likely that the threat actors responsible for the reported activities were seeking to deploy ransomware within the targeted networks. The steps which the threat actors took to make progress towards achieving this objective resulted in highly unusual patterns of network traffic. Darktrace’s detection of these unusual network activities allowed security teams to prevent these threat actors from achieving their disruptive objectives. 

Darktrace Coverage

Once threat actors succeeded in tricking users into running BumbleBee on their devices, Darktrace’s Self-Learning AI immediately detected the command-and-control (C2) activity generated by the loader. BumbleBee’s C2 activity caused the following Darktrace models to breach:

·      Anomalous Connection / Anomalous SSL without SNI to New External

·      Anomalous Connection / Suspicious Self-Signed SSL

·      Anomalous Connection / Rare External SSL Self-Signed

·      Compromise / Suspicious TLS Beaconing To Rare External

·      Compromise / Beacon to Young Endpoint

·      Compromise / Beaconing Activity To External Rare

·      Compromise / Sustained SSL or HTTP Increase

·      Compromise / Suspicious TLS Beaconing To Rare External

·      Compromise / SSL Beaconing to Rare Destination

·      Compromise / Large Number of Suspicious Successful Connections

·      Device / Multiple C2 Model Breaches 

BumbleBee’s delivery of Cobalt Strike Beacon onto target systems resulted in those systems communicating with Cobalt Strike C2 servers. Cobalt Strike Beacon’s C2 communications resulted in breaches of the following models: 

·      Compromise / Beaconing Activity To External Rare

·      Compromise / High Volume of Connections with Beacon Score

·      Compromise / Large Number of Suspicious Successful Connections

·      Compromise / Sustained SSL or HTTP Increase

·      Compromise / SSL or HTTP Beacon

·      Compromise / Slow Beaconing Activity To External Rare

·      Compromise / SSL Beaconing to Rare Destination 

The threat actors’ subsequent port scanning and SMB enumeration activities caused the following models to breach:

·      Device / Network Scan

·      Anomalous Connection / SMB Enumeration

·      Device / Possible SMB/NTLM Reconnaissance

·      Device / Suspicious Network Scan Activity  

The threat actors’ attempts to obtain account password data from domain controllers using the DCSync technique resulted in breaches of the following models: 

·      Compromise / Unusual SMB Session and DRS

·      Anomalous Connection / Anomalous DRSGetNCChanges Operation

Finally, the threat actors’ attempts to internally distribute and execute payloads resulted in breaches of the following models:

·      Compliance / SMB Drive Write

·      Device / Lateral Movement and C2 Activity

·      Device / SMB Lateral Movement

·      Device / Multiple Lateral Movement Model Breaches

·      Anomalous File / Internal / Unusual SMB Script Write

·      Anomalous File / Internal / Unusual Internal EXE File Transfer

·      Anomalous Connection / High Volume of New or Uncommon Service Control

If Darktrace/Network had been configured in the targeted environments, then it would have blocked BumbleBee’s C2 communications, which would have likely prevented the threat actors from delivering Cobalt Strike Beacon into the target networks. 

Figure 9: Attack timeline

Conclusion

Threat actors use loaders to smuggle more harmful payloads into target networks. Prior to March 2022, it was common to see threat actors using the BazarLoader loader to transfer their payloads into target environments. However, since the public disclosure of the Conti gang’s Jabber chat logs at the end of February, the cybersecurity world has witnessed a shift in tradecraft. Threat actors have seemingly transitioned from using BazarLoader to using a novel loader known as ‘BumbleBee’. Since BumbleBee first made an appearance in March 2022, a growing number of threat actors, in particular ransomware actors, have been observed using it.

It is likely that this trend will continue, which makes the detection of BumbleBee activity vital for the prevention of ransomware deployment within organisations’ networks. During April, Darktrace’s SOC team observed a particular pattern of threat actor activity involving the BumbleBee loader. After tricking users into running BumbleBee on their devices, threat actors were seen instructing BumbleBee to drop Cobalt Strike Beacon. Threat actors then leveraged Cobalt Strike Beacon to conduct network reconnaissance, obtain account password data from internal domain controllers, and distribute malicious payloads internally.  Darktrace’s detection of these activities prevented the threat actors from achieving their likely harmful objectives.  

Thanks to Ross Ellis for his contributions to this blog.

Appendices 

References 

[1] https://blog.google/threat-analysis-group/exposing-initial-access-broker-ties-conti/ 

[2] https://securityintelligence.com/posts/trickbot-gang-doubles-down-enterprise-infection/ 

[3] https://www.proofpoint.com/us/blog/threat-insight/bumblebee-is-still-transforming

[4] https://www.cynet.com/orion-threat-alert-flight-of-the-bumblebee/ 

[5] https://research.nccgroup.com/2022/04/29/adventures-in-the-land-of-bumblebee-a-new-malicious-loader/ 

[6] https://www.bleepingcomputer.com/news/security/conti-ransomwares-internal-chats-leaked-after-siding-with-russia/ 

[7] https://therecord.media/conti-leaks-the-panama-papers-of-ransomware/ 

[8] https://www.secureworks.com/blog/gold-ulrick-leaks-reveal-organizational-structure-and-relationships 

[9] https://www.prodaft.com/m/reports/Conti_TLPWHITE_v1.6_WVcSEtc.pdf 

[10] https://www.kroll.com/en/insights/publications/cyber/bumblebee-loader-linked-conti-used-in-quantum-locker-attacks 

[11] https://symantec-enterprise-blogs.security.com/blogs/threat-intelligence/bumblebee-loader-cybercrime 

[12] https://isc.sans.edu/diary/TA578+using+thread-hijacked+emails+to+push+ISO+files+for+Bumblebee+malware/28636 

[13] https://isc.sans.edu/diary/rss/28664 

[14] https://www.logpoint.com/wp-content/uploads/2022/05/buzz-of-the-bumblebee-a-new-malicious-loader-threat-report-no-3.pdf 

[15] https://ghoulsec.medium.com/mal-series-23-malware-loader-bumblebee-6ab3cf69d601 

[16]  https://blog.cyble.com/2022/06/07/bumblebee-loader-on-the-rise/  

[17]  https://asec.ahnlab.com/en/35460/ 

[18] https://thedfirreport.com/2021/07/19/icedid-and-cobalt-strike-vs-antivirus/

Inside the SOC
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Sam Lister
Specialist Security Researcher

More in this series

No items found.

Blog

/

/

July 7, 2026

Securing AI: Analysis of the Complete Security Stack with Governance and Controls

ai security stackDefault blog imageDefault blog image

Why traditional cybersecurity approaches are not enough for AI

AI adoption outpaces most security programs’ ability to adapt.  That gap is now one of the most consequential sources of cyber risk facing enterprises. As organizations embed generative and agentic AI into development workflows, business operations, and security tooling itself, the question is no longer whether AI will introduce risk. The question is whether organizations understand where that risk actually lives and how to manage it operationally.  

Two recent pieces of guidance underscore this shift:

  1. The upcoming Cybersecurity Framework Profile for AI from NIST
  1. The Five Eyes government guidance on the careful adoption of agentic AI services

Taken together, they point to a critical conclusion. AI security cannot be reduced to model hardening or prompt filtering. It requires a defense in depth strategy that treats AI as both a new attack surface and a force multiplier for defense, while accounting for how AI fundamentally changes scale, speed, and autonomy.  

Recent threat research suggests that today's cyber risk is driven less by initial compromise and more by an adversary's ability to blend into normal operations over time. AI systems create the same exposure in a new form: more autonomy, more scale, and more opportunities for risky behavior to blend into normal operations.

How NIST defines the three core pillars of AI security

The NIST profile organizes AI risk across three inseparable focus areas that span all cybersecurity functions, Secure, Defend and Thwart. These areas are not sequential. They exist simultaneously and must be addressed together.

Secure

This treats AI as an attack surface. It includes models, prompts, agents, pipelines, training and inference data, retrieval augmented generation corpora, and the AI supply chain itself. AI systems are opaque, probabilistic, and non-deterministic by design. Some vulnerabilities are inherent in how models are trained or how data is sourced. Traditional patching does not fully mitigate these risks. This is also where many enterprises are weakest today and, critically, where many security programs stop.  

Defend

This is AI as a defensive force multiplier. AI can improve detection speed, scale, correlation, and response, but only if the right models are used and operationalized correctly. Machine-speed behavior-based detection, response and containment becomes critical in defending non-deterministic systems. Accuracy, explainability, governance, testing, validation, and integration into SOC workflows matter as much as capability. Without those controls, hallucination risk, over automation, and misplaced trust become security risks themselves.  

Thwart

This treats AI as an adversarial accelerant. Threat actors are already using AI to generate targeted social engineering attacks, deepfakes, malware, and autonomous attack agents. Asymmetric warfare is highlighting faster vulnerability discovery and exploitation with a lag on patch development, testing and deployment.  

How this looks in practice

Darktrace researchers observed scaled, automated exploitation of the React2Shell vulnerability within days of disclosure. A vulnerable cloud asset was exploited in under 120 seconds of being deployed. Darktrace research team observed an AI/LLM-generated malware sample used in exploitation activity tied to React2Shell. The significance isn't novelty. It is that AI lowers the barrier to producing usable offensive tooling and compresses the time between experimentation and deployment.  

Tactics are getting more and more creative in order to string together steps of an attack kill chain. This creates a dependency on behavior-based detection, autonomous investigation, autonomous containment, training, resilience investment, and recovery planning across the entire enterprise.

Why agentic AI fundamentally changes enterprise cyber risk

The Five Eyes guidance on agentic AI highlights material changes to the cyber risk profile of an organization. Unlike generative AI systems that produce content for human consumption, agentic AI systems reason, plan, and act autonomously across tools, data, and environments. That autonomy, combined with access to real systems, amplifies the impact of traditional cyber failures and introduces new system level risks that are difficult to predict, observe, and contain.  

Risk in agentic systems does not live in the model alone. It emerges from interactions between models, prompts, memory, tools, APIs, identities, privileges, inter-agent trust relationships, and human assumptions baked into design. Vulnerabilities are often introduced through data, connectors, natural language interfaces, protocols, and drift by design.

In supply-chain incidents, attackers did not need sophisticated exploits to scale impact. They abused trusted systems built for automation and implicit access. Agentic AI inherits that model. Once a system can act across tools, data, and workflows, compromise propagates through trust relationships that were never designed for machine autonomy.

The major agentic AI risk classes include the following:  

  • The identity control for non-human identities or autonomous agents makes it difficult to mitigate over-permissioning, limiting access, scope, and duration, as well as access hygiene
  • Agents are frequently over permissioned
  • Compromised tools inherit agent authority
  • Static secrets enable impersonation
  • Implicit trust between agents enables lateral movement

Design and configuration risks compound this, including privileges evaluated once at startup, poor segmentation, unvetted third party tools, reused authorization decisions outside their original context, and guardrail limitations.  

Behavioral risk  

Agents can optimize for goals in unsafe ways, misinterpret ambiguous intent, chain actions into unintended sequences, change behavior during evaluation, and exhibit deceptive or sycophantic responses.  

Structural risk  

Structural risk follows from agentic systems that are tightly coupled, multicomponent ecosystems. Failures can propagate across agents. Hallucinations cascade downstream. Resource exhaustion becomes systemic. Tool misuse enables indirect prompt injection and command execution. Rogue agents can poison peer agents through trust relationships.  

Accountability

Accountability becomes unclear as autonomy increases. Autonomous agents assume human identity permissions, and humans should have clear ownership of these agents, but they don’t, and this model is flawed. Decision paths are opaque and non-deterministic. Logs are fragmented and difficult to interpret. Reproducing an incident will be impossible without explicit design for observability and forensics. An agent compromise is functionally an insider threat, often with better access and fewer behavioral constraints than a human.  

What does defense in depth look like for AI?

Agentic AI runs on software, networks, identities, and data. It must be governed using the same foundational principles that have proven resilient under uncertainty, including secure by design, defense in depth, zero trust, least privilege, continuous monitoring, behavior-based advanced threat detection and containment, and incident response and recovery.

Core components to a Defense in depth Strategy for Securing the use of AI:

  • Strong, precise identity control plane to include an identity per agent (cryptographic, non‑shared)
    • Privilege monitoring and just‑in‑time access
  • Data Governance
  • Secure‑by‑default configurations
    • Security Posture Management  
    • Zero Trust principles  
  • Strong guardrails, deny‑by‑default policies, and isolation
  • Explicit instruction hierarchies and controlled context
  • Behavioral-based detection across entire enterprise to include inputs, tools, and outputs as well as AI used on the endpoint, across the network, cloud, SaaS, email, and OT
    • Runtime anomaly detection and goal‑drift detection
    • Autonomous containment to mitigate risk and minimize damage
  • Hard boundaries on autonomy and delegation
  • Testing, Evaluation, Validation and Verification  
    • Determine when autonomous action and when human in the loop
    • Adversarial training and agent‑specific testing
    • Simulation, red teaming, and chaos testing
  • Kill‑switches, rollback, and containment mechanisms
    • Forensics data captures, interpretability, autonomous containment, and remediation/recovery plans  

Until standards, tooling, and assurance methods mature, organizations should assume agentic AI systems will behave unexpectedly and design deployments around resilience, behavior-based detection, reversibility, and containment, not efficiency.

How security leaders should prepare for enterprise AI adoption

AI security is not model security alone. Data, pipelines, identities, and agents are first class assets. Many AI attacks succeed through standard cyber failures amplified by AI. Identity, data, and supply chain risk dominate. Behavior-based detection and response are critical, not optional. Logging, provenance, versioning, and forensics data capture of detections are mandatory because you cannot investigate or recover from AI incidents without them.  

Risk will often be visible in behavior before it is clearly defined in policy or guidance. The same pattern has been seen in pre-CVE disclosure detection, where abnormal activity appears before the industry has named or described the vulnerability. AI systems introduce that uncertainty by design.

Security leaders should prioritize controls before AI is fully deployed, avoid generic AI security checklists, integrate AI risk into existing cyber programs, and mitigate the risk of non-deterministic technology with continuous oversight, monitoring, behavior analytics, anomaly detection, autonomous investigation, and autonomous containment.

Visibility has a different connotation with AI. Previously, audit logging worked for software/people, but with Generative AI-based systems, interpretability and explainability is difficult to understand, you cannot "undo" what has been done, or see the logic or control a chain of events. This is why behavioral-based detections and containment becomes critical.  

What capabilities should every AI security program include?

If an organization asked “what must be in place before scaling AI?”:

  1. AI Risk board and approval workflow
  1. IAM + PAM for all AI services and agents
  1. AI asset inventory
  1. Prompt/output DLP with sanctioned AI access – This is not just pre- and post- filters, but behavior-based detections of semantic interface as well as behavior-based analysis of output with associated risk context.  
  1. Shadow AI identification
  1. Secure MLOps – This is an entire paper itself
  1. Runtime guardrails and tool restrictions
    • Including AI Gateway/SASE/Zero trust/
  1. Runtime security with behavior-based detections
    • Complete visibility, monitoring, behavior analytics, anomaly detection, risk/intent/context evaluation of anomalies, autonomous investigation and autonomous containment of all AI assets across endpoint, network, SaaS, SASE, cloud, OT, email, and messaging platforms
  1. Secure data pipelines and data governance
  1. SOC workflow changes from malicious classification workflows to behavior-based detection workflows
  1. Remediation plans for AI-related incidents  

Layered Governance and Security Stack for Securing AI  

The following outline considers governance and security tools that should be considered, well-integrated, deployed, tested, operationalized and embedded within security workflows. These tools and controls map to NIST’s CMF for AI.  

These considerations do not need to be implemented in order. Runtime Detect and Respond will help mitigate risk while Governance, Visibility, and Identity mature.

Category Tooling Controls
Governance & Visibility
  • AI asset inventory / AI CMDB
  • Shadow AI discovery
  • SaaS discovery
  • AI usage on non-endpoint managed systems via network or cloud telemetry
  • MCP server/client usage via protocols
  • Browser telemetry
  • Gateway or SASE telemetry
  • Establish a risk board to set up controls
  • Mandatory registration of AI systems
  • Owner, data classification, intended use, and risk tier
  • Supplier disclosure requirements
  • Risk mitigation plan for AI adoption, innovation, or development
Identity, Access & Agent Control

Non-human autonomous agents should not have the full permissions associated with a human user.

  • IAM with workload identities
  • PAM for AI service accounts
  • Secrets management with short-lived tokens
  • Zero Trust principles
  • Identity, permission, and token hygiene
  • Unique identities per model, agent, and pipeline
  • Least privilege for tools, data, and APIs
  • Explicit approval for autonomous actions
Data Security & Privacy
  • Data classification and labeling
  • Enterprise DLP across endpoint, email, network, cloud, and SaaS
  • Forensics data capture after risky detections
  • Prompt-level DLP through behavior-based semantic analysis with risk and intent context
  • Input/interface analysis for risky data requests
  • Output analysis for sensitive data
  • Data integrity evaluation
  • Retention and redaction policies for prompts and responses
Secure MLOps / LLMOps
  • Secure CI/CD with AI-specific gates
  • Model registries with approval workflows
  • Dependency, container, and artifact scanning
  • SBOM/AIBOM generation
  • IaC security scanning
  • Security posture management
  • Misconfiguration identification
  • Hardening recommendations
  • Signed models and prompts
  • Versioned datasets, configurations, logging, and controls
  • Securing data pipelines
  • Controlled promotion
  • Quality assurance
  • Adversarial testing
Runtime Security

Securing runtime goes beyond guardrails and model firewalls to include behavior-based detections, response, and containment.

  • Detection, monitoring, and SOC integration
  • Centralized visibility into prompts, outputs, and tool calls
  • AI-specific detections
  • Behavior-based detection for AI usage patterns
  • Model drift and behavior monitoring
  • Autonomous containment
  • Behavior-based detection of model inputs and outputs
  • Prompt injection detection
  • Model manipulation, including jailbreaking, poisoning, and related attacks
  • Sensitive data access attempts
  • Behavior-based detection across low-code agents, high-code agents, MCP clients and servers, endpoint, network, cloud, email, SaaS, SASE, IoT, and OT
  • Policy enforcement between users, models, tools, agents, SaaS models/tools, and MCP servers/clients
  • Risk, intent, and context evaluation for detections and response actions
Response & Recovery
  • Autonomous containment
  • AI-assisted playbooks
  • Forensics data capture for AI-related events
  • Model rollback mechanisms
  • Backup and restore for models and datasets
  • Kill switch for agents
  • Autonomous response to agents performing risky behaviors
  • Model and dataset rollback
  • Remediation plans
  • Tabletop exercises
  • Supplier coordination plans
  • Post-incident AI performance validation

AI security requires continuous visibility and behavioral detection

AI changes how fast systems move, how decisions are made, and how risk propagates. It does not change the fundamentals of security. Organizations that succeed will be the ones that apply those fundamentals rigorously, assume failure, and build systems that can detect, contain, and recover when AI behaves in ways they did not anticipate. Security is not what AI is allowed to do. It is whether the organization can understand, trust, and control what AI actually does in practice.  

Take this guidance to understand different initiatives that organizations should be considering. Securing AI is the most critical component to AI safety. As organizations invest more in AI adoption, they should be investing in security in order to mitigate the risk of AI adoption. Organizations should be evaluating their governance and security stack to include well-integrated tools that are deployed, tested, operationalized and embedded within security workflows. While organizations mature in governance, visibility and identity access management, they should be investing in behavior-based detection and autonomous containment to mitigate AI risk.  

Continue reading
About the author

Blog

/

/

July 6, 2026

NIST Just Proved It: AI Security Can’t Be Solved With Rules

ai security nistDefault blog imageDefault blog image

Static AI guardrails are inherently limited

As organizations adopt generative AI, many still assume that the right set of guardrails will be enough. The problem is you can’t anticipate every way these systems might be misused, abused or attacked. What NIST has done is put a mathematical foundation under that intuition.

In recent research building on Gödel’s incompleteness theorems, which showed that any system built on a fixed set of rules will always have gaps, NIST demonstrates that there is no finite set of guardrails that can be universally robust against adversarial prompts. In plain terms, if your defense is based on a fixed set of rules, there will always be inputs that bypass them. Not because the rules are badly written, but because the problem space is bigger than static rules can ever cover.

This is not new in cybersecurity - detection rules have always had to live with this trade-off. What is different with GenAI is the scale and shape of that problem. These systems are built on human language, and human language is not bounded. It is fluid, contextual and deliberately ambiguous. The number of ways intent can be hidden is effectively limitless. You are not defending against a defined protocol or a fixed exploit chain. You are defending against the entire expressive capacity of people.

So attempting to create a complete set of rules is the wrong starting point. It assumes the problem can be deterministically described. NIST’s work shows that it cannot. Organizations still need a way to manage AI risk, but the traditional approach of defining allowed and disallowed patterns is always going to lag behind what is actually happening. The same input can be benign in one context and risky in another, and static rules struggle to capture that distinction.

The question then is what fills that gap?

AI security must shift from rules to behavior

What's required is a shift in what you are trying to understand. Rules try to describe what should and shouldn't happen. Behavior shows you what is happening. Or to put it another way, if inputs are unbounded and adversaries adapt, the only stable signal is behavior.

In a GenAI context, that means analyzing how an AI model is being used, how prompts evolve over time, how outputs are shaped, and where AI agent interactions start to drift from what is expected. It means moving from static definitions of bad to a more dynamic understanding of intent.

Instead of trying to predict every bad prompt, you focus on identifying when behavior starts to move outside expected norms. Instead of asking whether a single input matches a rule, you ask whether the overall pattern of activity makes sense for the system and how it’s being used.

Guardrails remain important but they are only one layer

This does not eliminate the need for guardrails. They still play a role. But they will never address the entire problem space and are simply one part of your defense in depth approach.

NIST’s proof is useful because it makes this explicit. It removes the assumption that with enough effort, a complete rule set is achievable. It isn’t.

Once you accept that, the shift becomes unavoidable. This is no longer a problem of writing better rules, but of understanding behavior in a space where the possible inputs are effectively unbounded.

For security leaders, that changes the nature of the problem. It is less about defining what should be allowed, and more about recognizing when something is no longer consistent with expected behavior.

That does not remove the need for guardrails, but it does change their role. They set boundaries, but they do not define understanding. The gap between the two is where risk now sits.

In the end, this is what “can’t be solved with rules” really means. Rules will always leave gaps, and those gaps are not theoretical. They show up in how systems actually behave Not what we expect them to do, or what we intended them to do, but what they are doing in practice. That is where the signal is, and increasingly, that is where the security problem sits.

References:

https://www.nist.gov/news-events/news/2026/06/nist-mathematical-proof-supports-transition-continuous-monitor-and-update

https://ieeexplore.ieee.org/document/11475847

Continue reading
About the author
Andrew Hollister
Principal Solutions Engineer, Cyber Technician
Your data. Our AI.
Elevate your network security with Darktrace AI