Threat reportThreat IntelligenceTL-2026-2113
Known Techniques, Unknown Speed: Aqua Security on How Frontier AI Collapses the Container Attack Chain
Aqua Security argues frontier AI models execute known container/cloud-native attack chains end-to-end without human direction at a speed that eliminates the discovery-to-exploitation window, citing Anthropic's Mythos model finding 271 previously unknown Firefox vulnerabilities as illustrative evidence. Aqua outlines a six-stage AI-driven attack pattern against containerized environments and argues simultaneous runtime enforcement across every stage is the required defense.
Key facts
| Identifiers | TL-2026-2113 |
|---|---|
| Also known as | Known Techniques Unknown Speed, AI-Driven Container Attack Chain, Autonomous Runtime Collapse |
| Severity | HIGH |
| Affected products | Industry-wide Containerized and Kubernetes-orchestrated workloads; Mozilla Firefox (illustrative Mythos exploit-discovery target); Meta / Vercel React Server Components / Next.js (CVE-2025-55182 'React2Shell', cited container initial-access precedent) |
| Malware | TeamPCP, VoidLink |
| Tools | Anthropic Claude 'Mythos' (frontier vulnerability-research model), Peirates - S0683, curl, wget |
| MITRE ATT&CK techniques | T1046 Network Service Discovery, T1059 Command and Scripting Interpreter, T1071 Application Layer Protocol, T1078 Valid Accounts, T1190 Exploit Public-Facing Application, T1496 Resource Hijacking, T1528 Steal Application Access Token, T1552 Unsecured Credentials, T1574 Hijack Execution Flow, T1595 Active Scanning, T1611 Escape to Host, T1613 Container and Resource Discovery |
| Motivation | UNKNOWN |
| Target sectors | technology, cloud service providers, financial services, cryptocurrency exchanges, critical infrastructure, government administration |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 8 cited sources |
Detection coverage
As of 2026-08-22, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2113 across Splunk SPL, Microsoft KQL and Sigma, covering 12 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: LD_PRELOAD Dynamic-Linker Hijack Into a Privileged Host Process Pre-Namespace-Isolation
Detects the execution/persistence stage of Aqua's chain: LD_PRELOAD environment-variable injection or a write to /etc/ld.so.preload targeting a privileged host process before namespace isolation completes, combined with container-breakout indicators (namespace change by an unprivileged program, container drift). Maps to the dynamic-linker-hijacking persistence technique and container-escape privilege escalation Aqua names explicitly.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1574, T1574.006, T1611
- Data sources: falco, auditd, sysmon_for_linux
- Severity: critical
- Confidence: high
Splunk SPL
index=falco sourcetype=falco:json rule_name IN ("Set Setuid or Setgid bit","Launch Privileged Container","Contact K8s API Server From Container","Write below etc")
| eval ld_preload_hit=if(match(rule_desc,"(?i)LD_PRELOAD|ld\.so\.preload") OR match(fd_name,"(?i)/etc/ld\.so\.preload"), 1, 0)
| where ld_preload_hit=1
| stats count earliest(_time) as first_seen latest(_time) as last_seen values(proc_cmdline) as commands by container_id, k8s_pod_name, k8s_namespace, proc_pname
| join type=left container_id
[ search index=falco sourcetype=falco:json rule_name IN ("Container Drift Detected","Namespace change (setns) by unprivileged program")
| stats count as breakout_events earliest(_time) as breakout_time by container_id ]
| table container_id k8s_pod_name k8s_namespace proc_pname commands first_seen last_seen breakout_events breakout_timeMicrosoft KQL
let LdPreloadInjection = DeviceProcessEvents
| where Timestamp > ago(24h)
| where ProcessCommandLine has "LD_PRELOAD"
| project InjectTime=Timestamp, DeviceId, AccountName, InitiatingProcessFileName, ProcessCommandLine;
let PreloadFileWrite = DeviceFileEvents
| where Timestamp > ago(24h)
| where FolderPath has "/etc/ld.so.preload" and ActionType in ("FileCreated","FileModified")
| project WriteTime=Timestamp, DeviceId, FileName, FolderPath, InitiatingProcessFileName;
LdPreloadInjection
| join kind=inner (PreloadFileWrite) on DeviceId
| where WriteTime between (InjectTime .. InjectTime + 5m)
| project DeviceId, AccountName, ProcessCommandLine, InjectTime, WriteTime, FolderPathSigma
title: Dynamic Linker Hijacking via LD_PRELOAD Environment Injection or /etc/ld.so.preload Write
id: b4e2a7f0-2c9d-4e11-9a3f-1a2b3c4d5e08
status: experimental
description: Detects LD_PRELOAD injection on the command line or a write to /etc/ld.so.preload, used to hijack execution flow of a privileged host process before namespace isolation completes.
references:
- https://www.aquasec.com/blog/known-techniques-unknown-speed-how-ai-changes-the-attack-chain/
author: Threadlinqs Team
date: 2026/08/22
tags:
- attack.persistence
- attack.t1574
- attack.t1574.006
- attack.privilege-escalation
- attack.t1611
logsource:
category: process_creation
product: linux
detection:
selection:
CommandLine|contains: 'LD_PRELOAD='
condition: selection
falsepositives:
- Legitimate use of LD_PRELOAD by APM/observability agent shims (Datadog, New Relic profiling agents)
level: criticalKnown false positives:
- Legitimate use of LD_PRELOAD by APM/observability agent shims (Datadog, New Relic profiling agents)
- Custom memory allocators (jemalloc, tcmalloc) deployed via LD_PRELOAD in performance-tuned workloads
Tuning notes: Allow-list known observability-agent and allocator hashes/paths that legitimately set LD_PRELOAD; require the breakout_events join (Container Drift / setns by unprivileged program) before alerting at critical severity in noisy environments.
- Kubernetes Audit Trail: Container/Resource Discovery Immediately Preceding Sidecar Proxy Object Mutation · unlock with TL Creds
- Reverse Shell Process Chain Combined With Kernel-Primitive Container Escape Indicators · unlock with TL Creds
- Kubernetes Service-Account Token Theft From Default Mount Path Followed by Cross-Source API Server Authentication · unlock with TL Creds
- Web-Facing Process Spawning Interactive Shell Immediately After Crash/Panic (Pre-Auth RCE-to-Shell) · unlock with TL Creds
- Fileless Payload Execution via memfd_create/tmpfs Preceding Namespace Manipulation · unlock with TL Creds
- Post-Compromise Impact Triad: Cryptominer Signature, Mounted-Secrets Exfiltration, or Mass PV Encryption · unlock with TL Creds
- Container Recon Burst Preceding Sidecar/Public-Facing Application Exploitation · unlock with TL Creds
- TLS C2 Beaconing via curl/wget With Regular Sleep/Jitter Timing to Non-Standard Destinations · unlock with TL Creds
Indicators of compromise
As of 2026-08-22, Threadlinqs Intelligence tracks 12 indicators of compromise for TL-2026-2113.
Technical analysis
This is a TREND / threat-landscape item, not a single-CVE threat. Aqua Security (author: Erin Stephan, Head of Product Marketing; published May 7, 2026, category 'Autonomous Runtime Security') argues in 'Known Techniques, Unknown Speed: How AI Changes the Attack Chain' that frontier AI models introduce no fundamentally novel attack techniques against containerized/cloud-native environments, but instead execute the same well-documented technique chains end-to-end, autonomously, and at a speed that collapses the traditional discovery-to-exploitation window that defenders rely on for triage and patching.
Aqua's central evidence point is real and independently corroborated: Anthropic's 'Mythos' model (a Claude-family frontier vulnerability-research system, unveiled April 2026) was used by Mozilla to analyze Firefox and surfaced 271 previously unknown vulnerabilities, which shipped fixed in Firefox 150 on April 22, 2026. Three received public CVE credit to the model (CVE-2026-6746, CVE-2026-6757, CVE-2026-6758); most of the 271 were defense-in-depth issues, hardening flaws, or bugs in non-exploitable code paths that did not individually meet the bar for a CVE. Firefox CTO Bobby Holley characterized the result as significant in scale but not in kind: 'we also haven't seen any bugs that couldn't have been found by an elite human researcher,' rejecting the idea that Mythos found 'entirely new forms of vulnerabilities.' Aqua's blog frames the same result as '271 previously unknown exploit vectors in Firefox in under 72 hours' against a Mozilla estimate of four to seven months of manual red-teaming for equivalent coverage; neither SecurityWeek's nor TechCrunch's independent reporting on the disclosure specifies an exact discovery timeframe, so the 72-hour figure should be read as Aqua's own framing of the finding rather than an independently confirmed measurement. SecurityWeek separately reports Palo Alto Networks testing showing Mythos achieving 'the equivalent of a year's worth of pentesting in under three weeks' on general assessments (not Firefox-specific).
Aqua's illustrative six-stage AI-driven container attack chain: (1) Reconnaissance/Resource Development — kernel-version enumeration, mounted-volume probing, and cluster metadata-endpoint querying, plus parallel port scanning (5+ privileged and 10+ ephemeral ports within 5 minutes); (2) Delivery/Initial Access — exploitation of a sidecar container logic flaw, including novel vulnerability analysis that does not require a known CVE; (3) Execution/Persistence — reverse-shell spawning, LD_PRELOAD shared-library injection into an existing privileged host process before namespace isolation completes, and fileless payload deployment into writable-executable memory; (4) Command and Control — outbound TLS connections to attacker infrastructure using standard pre-installed tools (curl, wget) for staging and persistence; (5) Lateral Movement/Privilege Escalation — kernel-primitive exploitation for container breakout combined with theft and reuse of the Kubernetes service-account token mounted inside the container to authenticate to the API server; (6) Impact — cryptominer deployment, data exfiltration from mounted secrets volumes, and ransomware distribution across persistent volumes. Aqua names no specific CVE, malware sample, or threat actor for this chain — it is presented generically — and argues that because an autonomous agent moves through all six stages without pausing for human decision points, only runtime enforcement operating simultaneously at every stage (not point-in-time scanning or patch-velocity programs) can interrupt it before impact.
Every individual technique in Aqua's generic chain has a documented real-world precedent that grounds the analysis: Unit 42 (Palo Alto Networks) reports Kubernetes service-account-token theft increased 282% year-over-year and was observed in 22% of cloud environments in 2025, with tokens mounted by default at /var/run/secrets/kubernetes.io/serviceaccount/token; malware frameworks TeamPCP and VoidLink now ship dedicated Kubernetes credential-harvesting modules; the Go-based post-exploitation tool Peirates (MITRE ATT&CK software S0683), used by the cloud-focused intrusion sets SCARLETEEL and TeamTNT, automates namespace/role enumeration, secret theft, cloud API misuse, and node attacks. Unit 42 also documents a mid-2025 cryptocurrency-exchange compromise attributed to Slow Pisces (Lazarus/TraderTraitor) that followed spearphishing → malicious pod deployment → service-account-token extraction → API enumeration → cloud-infrastructure compromise, and a distinct initial-access precedent for 'sidecar/application logic flaw' exploitation: CVE-2025-55182 ('React2Shell'), a CVSS 10.0 pre-authentication remote-code-execution vulnerability in React Server Components (react-server-dom-webpack/turbopack/parcel 19.0.0-19.2.0; Next.js 15.0.0-16.0.7) disclosed December 3, 2025 and observed under active exploitation against containerized application backends within two days, with attackers enumerating cluster resources and harvesting mounted service-account tokens post-exploitation — the exact recon-then-token-theft pattern Aqua's chain describes.
Aqua's prescriptive thesis is runtime enforcement that operates at every stage of the chain simultaneously, blocking a process from advancing to the next stage the moment it deviates from policy, rather than relying on triage workflows (SIEM alert → tier-1 pickup → ticket → patch) whose latency assumes a human-speed attacker.
MITRE ATT&CK techniques by tactic:
Discovery
T1046 Network Service Discovery; T1613 Container and Resource Discovery
Execution
T1059 Command and Scripting Interpreter
Command and Control
T1071 Application Layer Protocol
Privilege Escalation
T1078 Valid Accounts; T1611 Escape to Host
Initial Access
T1190 Exploit Public-Facing Application
Impact
Credential Access
T1528 Steal Application Access Token; T1552 Unsecured Credentials
stealth
Reconnaissance
Remediation
Patches
- Patch container-adjacent pre-auth RCE vectors on an emergency cadence — CVE-2025-55182 (React2Shell) was under active exploitation within two days of disclosure
- Track vendor advisories for vulnerability classes disclosed at AI-assisted speed (e.g., Firefox 150 / CVE-2026-6746, CVE-2026-6757, CVE-2026-6758) and prioritize triage accordingly
Immediate actions
- Enforce runtime policy simultaneously at every stage of the container lifecycle (reconnaissance through impact) rather than relying on point-in-time scanning, per Aqua's simultaneous-enforcement model
- Disable automatic service-account token mounting (automountServiceAccountToken: false) on workloads that do not require Kubernetes API server access
- Apply seccomp/AppArmor/SELinux profiles that block LD_PRELOAD environment-variable injection and dynamic-linker hijacking in running containers
Workarounds
- Restrict node-level kernel-primitive access (seccomp default profile, drop unneeded Linux capabilities) to reduce container-breakout surface
- Segment Kubernetes clusters/namespaces so a single sidecar compromise cannot reach cluster-wide metadata endpoints or cloud credentials
Longer-term hardening
- Adopt bound, time-limited, audience-restricted Kubernetes service-account tokens via the TokenRequest API instead of long-lived tokens mounted at /var/run/secrets/kubernetes.io/serviceaccount/token
- Deploy runtime detection for anomalous outbound TLS connections and curl/wget invocations originating from workload containers (LOLBin-style C2 staging)
- Enforce least-privilege RBAC on service accounts to limit blast radius if a token is stolen, per Unit 42's Slow Pisces cryptocurrency-exchange case study
- Harden sidecar-container trust boundaries and validate inter-container logic paths that can bypass namespace isolation checks
Affected and fixed versions
- Industry-wide — Containerized and Kubernetes-orchestrated workloads
Vulnerable versions: environments lacking runtime enforcement across every attack stage
Fixed in: environments enforcing continuous runtime policy from reconnaissance through impact - Mozilla — Firefox (illustrative Mythos exploit-discovery target)
Vulnerable versions: Firefox builds prior to 150
Fixed in: Firefox 150 (April 22, 2026) - Meta / Vercel — React Server Components / Next.js (CVE-2025-55182 'React2Shell', cited container initial-access precedent)
Vulnerable versions: react-server-dom-webpack/turbopack/parcel 19.0.0-19.2.0; Next.js 15.0.0-16.0.7 and specific canary builds
Fixed in: patched releases per vendor advisory, December 2025
Timeline
- Unit 42 documents a cryptocurrency-exchange compromise (approximate mid-2025 timing) attributed to Slow Pisces (Lazarus/TraderTraitor): spearphishing, malicious pod deployment, Kubernetes service-account-token theft, API enumeration, and cloud-infrastructure compromise.
- CVE-2025-55182 ('React2Shell'), a CVSS 10.0 pre-authentication RCE in React Server Components, is disclosed — the container-application initial-access precedent later echoed in Aqua's generic sidecar-exploitation stage.
- Unit 42 observes active exploitation of CVE-2025-55182 against containerized application backends within two days of disclosure, with attackers enumerating cluster resources and harvesting mounted Kubernetes service-account tokens post-exploitation.
- Palo Alto Networks testing, reported by SecurityWeek, shows Anthropic's Mythos model achieving 'the equivalent of a year's worth of pentesting in under three weeks' on general assessments.
- Slashdot reports Mozilla used Anthropic's Mythos model to find and fix 271 bugs in Firefox.
- Mozilla ships Firefox 150, patching all 271 vulnerabilities Mythos identified; three receive public CVE credit (CVE-2026-6746, CVE-2026-6757, CVE-2026-6758). Firefox CTO Bobby Holley states none of the bugs 'couldn't have been found by an elite human researcher.'
- TechCrunch publishes a feature on Mythos as an agentic vulnerability-research system that assesses and filters its own findings, including sandbox-escape bugs requiring multi-step exploit construction, and reports Firefox shipped 423 bug fixes in April 2026 versus 31 the prior year.
- Aqua Security (Erin Stephan) publishes 'Known Techniques, Unknown Speed: How AI Changes the Attack Chain,' citing the Mythos/Firefox result ('271 previously unknown exploit vectors in under 72 hours' vs. a 4-7 month manual-red-team estimate) as evidence for a six-stage AI-driven container attack chain and a simultaneous-runtime-enforcement defense thesis.
Sources cited
- NVD - CVE-2025-55182 Detail NVD (NIST) · Advisory · 2025-12-03
- Understanding Current Threats to Kubernetes Environments Unit 42 · Research · 2026-01-01
- Known Techniques, Unknown Speed: How AI Changes the Attack Chain aquasec.com · Blog · 2026-05-07
- Claude Mythos Finds 271 Firefox Vulnerabilities SecurityWeek · News · 2026-04-22
- Mythos found 271 Firefox flaws – none a human couldn't spot The Register · News · 2026-04-22
- How Anthropic's Mythos has rewritten Firefox's approach to cybersecurity techcrunch.com · News · 2026-05-07
- Mozilla Uses Anthropic's Mythos To Fix 271 Bugs In Firefox news.slashdot.org · News · 2026-04-21
- Anthropic Mythos finds 271 Firefox vulnerabilities aibusinessreview.org · News · 2026-05-09
Previous and next in threat intel
Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.