ARTEX is a free, open-source framework that lets one operator point a large language model at a target and have AI agents run the intrusion. Since late summer it has turned up in a supply chain compromise that reached more than 2,250 Microsoft 365 tenants, in bank breaches in South Korea, and in an espionage campaign across 14 countries.
The FBI has now issued a flash alert. This article covers what ARTEX does, how to spot it, and what to change, so it finds less to work with.
What ARTEX is and how attackers are using it
ARTEX appeared on GitHub in mid-2026 as a penetration testing tool. It is written in Go, works with almost any AI model, and divides the work among agents: planners map the target and choose the next step, workers run the tools, and verification agents check the results, with everything recorded in a shared graph of discovered assets and weaknesses (Microsoft, Security Affairs). The operator sets an objective, and the agents handle reconnaissance, exploitation, and data theft with little further input. The tool is free, and Unit 42 notes that an attacker needs little beyond access to a language model.
The developer moved the project to closed source on October 8 after it was abused, but forks, localized builds, archived copies, and a commercial version remain in circulation, according to the FBI alert as reported by The Record. Censys counted 1,465 exposed ARTEX consoles on November 3 and expects the real number to be higher, since many run behind firewalls.
Speed is what changes the defender's math. The FBI reports intrusions that went from first targeting to privileged access in under three hours. Unit 42 saw attackers reach cloud access in less than a day, in some cases without exploiting any software vulnerability, by using public password reset pages and exposed APIs. Here is what has been reported over the past several weeks.
| Campaign | Reported activity | Attribution |
ConnectWise Cloud Backup compromise Oct 10 to 16 | Stolen ConnectWise application credentials were used to attempt sign-ins to more than 2,250 Microsoft 365 tenants. In some tenants a ConnectWise app assigned itself Global Administrator, and agents read Azure Key Vault secrets in bulk. (Microsoft, Huntress) | Storm-0940, China-state aligned (Microsoft assessment) |
South Korean and Japanese financial institutions Late September onward | Agents found SQL injection and authentication bypass flaws in peripheral web apps such as loan inquiry portals and employee mobile backends. Zscaler reports personal data on about 67,000 people was taken from South Korean firms. The FBI also cites Japanese institutions. (Zscaler, CrowdStrike, The Record) | Unnamed, likely Chinese-speaking, financially motivated actor (moderate confidence) |
Vapor Panda From Oct 17 | More than 200 organizations in at least 14 countries were targeted over 11 days, with post-exploitation confirmed at nine, across government, telecom, pharmaceutical, semiconductor, defense, and financial sectors. Operators told the agents to skip Windows hosts running CrowdStrike Falcon and focus on unprotected ones. (BleepingComputer) | Vapor Panda, tracked by CrowdStrike |
AI credential theft and exchange KYC bypass Oct 3 to Nov 5 | Agents harvested exposed AI provider keys and probed AI gateways and development platforms for default credentials and known flaws. Separate sessions built a real-time deepfake pipeline against identity verification at two cryptocurrency exchanges. The vendor's server rejected every attempt. (Security Affairs) | Several China-linked groups (Amazon, moderate confidence) |
IT and software vendors October | Password spray, OAuth manipulation, and account recovery abuse against Okta, Entra ID, and AWS Cognito, aimed at software engineers. The median gap between attempts was one second, and most attempts failed. (Wiz) | Overlaps with Storm-0940 (Wiz) |
Software, industrial automation, semiconductor, and telecom targets Since late August | Agents ran SQL commands through an internet-facing Weaver OA server and used MySQL logging to place JSP web shells. Command output left through rotating webhook.site endpoints. (Google via GBHackers) | UNC7228 (Google Threat Intelligence Group). Overlap with Storm-0940 noted, not confirmed |
AWS environments October | Compromised long-term IAM keys and public Cognito user pools were the entry points, and every session began with unauthenticated S3 reads. The Cognito attempts failed because the role lacked permissions. (Datadog) | Not attributed |
Financial organizations (Storm-3358) From October | ARTEX was paired with conventional automated tools to scan internet-facing systems, exploit anonymous file upload and server-side request forgery (SSRF) flaws, and install open-source remote access trojans (WISP, vshell, and AdaptixC2). (Microsoft, Nov 6 report) | Storm-3358, financially motivated, Cambodia-based (Microsoft assessment) |
Two things stand out:
Amazon said none of the techniques in these campaigns were new, which matches what the other vendors describe: password spray, exposed secrets, default credentials, and overpermissioned apps.
And attribution is still thin. Microsoft's Storm-0940 assessment is the firmest, while the rest are moderate confidence or unnamed.
The attack chain, from the operator's seat
Microsoft Threat Intelligence reported on November 6 that Storm-3358 began aiming ARTEX at financial organizations in October. Microsoft describes the group as financially motivated and Cambodia-based, focused mainly on financial firms in Southeast Asia plus a smaller number of large financial organizations elsewhere. Nothing in its toolkit is new. The change is the time between steps. An attacker using ARTEX still needs an exposed upload endpoint or an SSRF-prone application. What ARTEX removes is the pause between finding that weakness and acting on it.
Every step below is something a competent tester has done by hand for years. Here is the chain, and what a human operator would normally have to do at each stage.
Find the entry point
Scan internet-facing hosts and applications for exposed services and weak endpoints.
Get code running
An upload endpoint that accepts files without authentication is often a short path to a web shell. SSRF lets an attacker make the server send requests on its behalf, which can reach internal services the internet cannot.
Bring in tools
Microsoft observed ARTEX agents using netcat and curl to download RATs onto compromised hosts.
Stay and spread
Storm-3358 has used the WISP, vshell, and AdaptixC2 RATs for persistence and lateral movement.
Turn access into money
In broader ARTEX activity, Microsoft reports agents compromising credentials from public sources and compromised systems, then using compromised cloud identities for persistence and data exfiltration. That is where the payoff sits for a financially motivated group. The second and third steps are the cheapest places for defenders to break the chain.
How an agent differs from a human operator
Agentic frameworks send a wide variety of requests in a short time, retry variations of the same technique, and move to next-stage actions at machine speed. A human tester prioritizes, gets bored, and sleeps. An agent keeps iterating. That changes three things for defenders.
Persistence of attempts
A human might try a handful of payloads against an upload form and move on. An agent keeps generating variations until something works or its budget runs out.
No pause between stages
Alert triage that assumes hours between initial access and lateral movement is calibrated for human attackers. Microsoft's separate November 5 research described an agent going from scanning GitHub for secrets to exfiltrating data from an Azure Key Vault in about five minutes (Dark Reading, secondary coverage). That example is not from the Storm-3358 report.
A lower skill floor
ARTEX agents can also brute-force passwords, spray credentials, enumerate and abuse APIs, and dump databases. A mid-tier financially motivated group gets parallel operations and persistence without growing its team.
By default, ARTEX uses Nmap, curl, wget, Python HTTP clients, Norma SDK WebFetch, and headless Chromium through Playwright, which can leave recognizable traffic such as default user agent strings (Microsoft). Blocking on those characteristics will not stop an agent, which can change them quickly, but spotting them can alert defenders. Treat a user agent block as an alert signal, not a control.
The targeting is concentrated in Southeast Asia, and the entry techniques are not regional: they apply to any organization with exposed web applications. The reporting also has limits. Microsoft does not name victims, say how many organizations were targeted or affected, report losses, or say how Storm-3358 obtained ARTEX.
Detecting and alerting on ARTEX AI attacks
ARTEX does not install anything on the victim, so much of the evidence sits in identity, cloud audit, and web logs rather than on endpoints (ThreatDefence). The tools its agents drop, such as web shells, are still catchable by EDR. What gives the agent away is its rhythm: a burst of requests about a second apart, a pause of 25 to 31 seconds while the model decides what to do, then another burst (Unit 42). One trait alone could be a noisy scanner or a human tester. Several together point to an agent. Build alerts on behavior first. Use the indicators below to sweep logs backward, because infrastructure rotates within minutes and user agents are easy to change.
Hunt back to at least September 1, 2026 (Sophos). For ConnectWise-related service principal sign-ins, Microsoft advises reviewing from October 15 at a minimum and from October 1 where records exist.
Behavioral detections
| Signal | What to look for | Where to look |
| Machine-paced password spray | Failed sign-ins across many users from one source, about one second apart, in waves of 20 to 60 minutes. A success from the same source after many failures is the higher-severity version. ThreatDefence's example thresholds: 200+ attempts, 20+ users, median gap of 2 seconds or less. (Wiz, ThreatDefence) | Entra sign-in logs (interactive and non-interactive), Okta, WAF |
| Think-then-act request cadence | Bursts of 1 to 14 requests about a second apart, then 25 to 31 seconds of quiet. A pivot within about a minute of obtaining a credential or key, and a follow-up GET to the same path after almost every successful write. (Unit 42) | Web, API gateway, storage access logs |
| Attack-map reconnaissance | Requests for source maps, /assets/config.json, /env.js, and OpenAPI or Swagger documents, followed quickly by calls to undocumented routes. A request for a path that cannot exist, such as /zzz-nonexistent-xyz123, before fuzzing starts. (Wiz, Datadog) | WAF, reverse proxy, CDN logs |
| Token and OAuth tampering | Any request carrying an unsigned JWT (header beginning eyJhbGciOiJub25lIn0.), OAuth callbacks with junk state such as callback?code=AAAA&state=notvalidb64!!, and many SSO route variants in minutes (more than 60 in two minutes at one victim). (Wiz, ThreatDefence) | WAF, SSO logs |
| Account recovery abuse and canary inputs | Repeated password reset, username validation, and activation calls. Bogus usernames tried before real ones, and placeholder values such as evil[.]com, test@example[.]com, and deadbeefdeadbeefdeadbeef12. (Unit 42, Wiz) | Identity provider and application logs |
| Application identity abuse | An application adding credentials to another application, an application assigning itself a directory role (Global Administrator and Exchange Administrator were both seen), or a vendor app signing in from an IP it has never used. (Microsoft, Huntress, Sophos) | Entra audit, service principal sign-in, and Graph activity logs |
| Cloud secret and storage probing | 20 or more successful secret reads across 3 or more Key Vaults in 15 minutes, AccessDenied bursts, write-then-delete tests on storage buckets, and first use of a cloud key from a new IP. In AWS: anonymous S3 GetObject and ListObjects calls at the start of a session, CloudWatch FilterLogEvents with a limit of 1000, and bursts of Cognito SignUp calls. (Microsoft, Unit 42, Datadog) | Key Vault diagnostics, CloudTrail, storage logs |
| Web shells and server-side execution | New .jsp or .aspx files in unexpected web application paths, SQL commands issued through application APIs, and MySQL logging used to write files into web-accessible directories. Bursts of commands and repeated variants of the same command. (Zscaler, Google via GBHackers, Sophos) | EDR, web server, database logs |
| Unusual egress | Server-side calls to model APIs such as api.deepseek.com, open.bigmodel.cn, api.x.ai, and api.anthropic.com. These are legitimate services, so investigate them only alongside other signals (Zscaler). Also watch for command output sent to webhook.site-style capture services (Google via GBHackers). | Proxy, DNS, firewall |
Storm-3358 adds two behaviors worth confirming you alert on, because they follow from the entry techniques Microsoft describes: a web server process that spawns a shell, and an internet-facing server making outbound curl or wget calls to hosts it does not normally contact.
Microsoft Defender customers should look for alerts titled "Application self-assigned Global Administrator role" and "Unusual Microsoft Graph activity associated with Compromised ConnectWise Application," plus the Key Vault anomaly alert KV_AccountVolumeAccessDeniedAnomaly. Microsoft's report includes Defender XDR and Sentinel hunting queries, and ThreatDefence has published example KQL for spray detection, application self-changes, and Key Vault reads.
Indicators of compromise
All values are defanged. Validate before blocking: several IPs sit on shared hosting and rotate. Amazon has published additional IP addresses for the AI credential theft and exchange KYC campaigns (Security Affairs summarizes them), and those should be pulled from Amazon's original report.
Storm-3358 (Microsoft, Nov 6 report): Microsoft's Sentinel hunting query lists 178.15.54[.]129, while its IOC table lists 178.16.54[.]129 as an ARTEX server.
Check both against Microsoft's post before loading either. Microsoft also notes that actors regularly use proxies and VPNs to change source IPs, so these indicators will age quickly.
Network infrastructure
| Indicator | Context | Source |
144.172.64[.]128 144.172.64[.]99 144.172.64[.]60 | Storm-0940. .128: app sign-ins and Graph activity in customer tenants. .99: password spray. .60: listed IOC. | Microsoft, Huntress |
| 144.172.94[.]243 | Broad failed login spray, no app sign-ins seen | Huntress |
187.15.88[.]186 187.15.89[.]186 187.15.90[.]217 | VPN addresses used for Graph API access through compromised apps, Oct 15 to 16 | Microsoft |
104.194.150[.]21 167.88.172[.]97 | IT and software vendor attacks | Wiz |
| 38.244.50[.]120 (port 18899) | ARTEX host used against South Korean financial organizations; open directory | CrowdStrike |
101.53.80[.]20 205.214.59[.]31 124.155.252[.]63 154.201.79[.]246 23.248.249[.]90 23.158.220[.]98 103.248.148[.]84 203.160.133[.]172 209.209.85[.]38 | Proxy layer in the same campaign | CrowdStrike |
xcai[.]pro hxxp://38.244.50[.]120:18899/.claude/CLAUDE.md | Likely LLM API proxy or reseller; open directory holding an attacker prompt file | CrowdStrike |
154.64.252[.]77 154.36.136[.]197 154.36.136[.]23 | NetLab Global hosting. .77 tied to a Cognito sign-up burst Oct 26 to 27. | Datadog |
103.54.154[.]40 103.54.154[.]32 | Zeofast Network hosting. .40 searched CloudWatch on Oct 20. | Datadog |
169.128.55[.]214 82.158.88[.]233 45.145.228[.]25 | Hosting providers | Datadog |
102.216.154[.]102 88.229.22[.]232 | ISP addresses (Kenya, Turkey) | Datadog |
45.153.126[.]223 (nginxmysql[.]com) 89.46.234[.]50 (sqlserver-sqlite[.]com) 192.121.47[.]20 (www[.]ddduuuppp[.]com) | Vapor Panda command and control | NeoTeo, SOCRadar (secondary; validate) |
Request artifacts and user agents
| Indicator | Context | Source |
artex-readonly/1.0 artex-pentest/1.0 artex-audit/1.0 artex-verify/1.0 artex-spawn/1 ARTEX-worker/1.0 artex-i21861-canary Mozilla/5.0 ARTeX Mozilla/5.0 (Windows NT 10.0; Win64; x64) ARTeX-authorized-pentest | User agents that name the tool. Easy for an operator to change, so absence proves little. | Datadog |
ReconProbe/1.0 norma/0.4 azure-toolkit/1.0 | Distinctive, but do not name ARTEX | Wiz, ThreatDefence |
Python-urllib/3.12 python-requests/2.31.0 python-requests/2.34.2 curl/7.88.1 | Common clients. Useful only combined with an IP, rhythm, or target match. | Microsoft, Huntress, Wiz |
/zzz-nonexistent-xyz123 zzz_nonexistent_abc.txt Notexist-xyz | Not-found canaries (web path and S3 keys) | Wiz, Datadog |
/packs/js/main-<hash>.js.map /assets/config.json /env.js /swagger/?format=openapi | Recon requests for source maps, runtime config, and API docs | ThreatDefence |
eyJhbGciOiJub25lIn0. callback?code=AAAA&state=notvalidb64!! | Unsigned token prefix and malformed OAuth callback | Wiz, ThreatDefence |
evil[.]com test@example[.]com deadbeefdeadbeefdeadbeef12 artex-hqheizfk[.]invalid | Canary and placeholder values in sign-up and login attempts | Unit 42, ThreatDefence |
Port 8787 (also 8790, 443, 80) Page title starting "ARTEX" with _next/static assets Renamed builds: ScopeWeaver | Console fingerprints, useful for checking your own exposure or finding an unexpected console inside your network | Censys |
/tasks /exploration/findings /exploration/graph /exploration/intents /llm/profiles /llm/pool /intercept | ARTEX console API paths | ThreatDefence |
Host and file artifacts
| Indicator | Context | Source |
Files or paths starting with artx, especially under C:\Windows\Temp Temp tooling named i<intentID>[variant]_<purpose>.<ext>, for example i12345_b2.ps1 Command markers such as MARKER123 and ARTX123 | Sophos presents the markers as examples, so treat them as patterns rather than fixed strings. | Sophos |
C:\ProgramData\artx5v\ D:\ProjectX\Merman_New\ Command comments prefixed art18- Accounts or files named artex-probe-* or ArtexProbe | Reported in Vapor Panda and Google/Mandiant cases | SOCRadar (secondary; validate) |
svchost.exe (renamed GameHook): 49fe70419fd19b81ca8d730a6ef7c1ec71b681d60b390761070d42686c7d2068 graphics-hook-filter64.dll: 3fb5a9fb3e2cb1d6208966bd4e55d10282d19c7250d827efa560463d60a12ee4 graphics-hook_config.bin: 364251838edfcf5ef1107fa692da3c5bd9a540482449f645992830134bf807ab | BLOODALCHEMY backdoor chain used by Vapor Panda. The renamed binary loads the DLL, which decrypts the config file. Hash algorithm is not labeled in the source; 64 characters is consistent with SHA-256. | NeoTeo, SOCRadar (secondary; validate) |
| Unexpected .jsp or .aspx files in web application paths | Possible web shells | Zscaler, Google via GBHackers |
| Sophos detections ATK/Artex-A, -B, -C, and Troj/Bckdr-SDA | Behavior detections and a trojanized ARTEX Docker image. Treat unofficial ARTEX builds as untrusted. | Sophos |
Microsoft 365 application IDs (ConnectWise)
| Application ID | Application |
| 506e08f9-6e9a-4da4-942b-a099a03644ae | ConnectWise Cloud Backup (formerly SkyKick) |
| 24205314-2b89-4e07-a2a4-579e898a5640 | Cloud Management, MS Teams (added itself to Global Administrator in one tenant) |
| 0de14c77-b9b0-4137-9360-34878a929827 | Cloud Management, MS Graph |
| b78f6a77-d4e5-46fb-87e5-d1648bd02519 | Cloud Management, Exchange Online (abuse reported Oct 17) |
| b2b55f5b-7758-4719-8d8f-88ae172e6bdb | Cloud Management, SharePoint Online (abuse reported Oct 17) |
| 2399d300-6fca-4e6a-a686-399fefe75b83 | Cloud Management |
| 412dcf96-176d-4b3a-a1b0-4981ad66971e | Cloud Management, Partner Center |
App IDs are from Huntress and Microsoft. Older tenants may show the apps as "SkyKick Cloud Manager." If your tenant uses ConnectWise Cloud Backup or Cloud Management, search by application ID rather than display name, and disable the apps instead of deleting them, since deletion removes consent and evidence.
Do not re-enable them until ConnectWise confirms credentials were reset. Review service principal sign-ins from non-Microsoft IPs, and review Entra audit logs for "Add member to role," "Update application," and "Add service principal credentials" events (Huntress). A clean sign-in result does not rule out access, because some Graph directory reads leave no audit record.
Practical adivce for long-term protection
Agents work from what you publish
Unit 42 and ThreatDefence both recommend removing source maps, OpenAPI specs, and config files that hold tokens or client IDs from production, and requiring authentication on every write endpoint, including development and staging APIs. Peripheral systems deserve the same attention as core ones. The FBI says ARTEX favors partner and supplier portals, mobile backends, employee applications, and legacy services, and the Korean victims described by Zscaler were a loan inquiry portal and a mobile work-support backend, not core banking systems.
Identity is the control that has to hold
Require phishing-resistant MFA for administrators and engineers, turn on smart lockout, block legacy authentication, and rate limit sign-in and password reset endpoints by account as well as by IP address, since agents rotate addresses when blocked (Wiz). Block local passwords for federated workforce accounts, make sure a reset cannot unlink a federated identity, reject unsigned tokens and SAML assertions, and validate OAuth state on every callback. Replace long-lived cloud keys with roles, and keep secrets in a secrets manager rather than in function configs or provisioning records (Unit 42, SOCRadar).
The ConnectWise incident was a vendor problem that became a tenant problem, so third-party applications deserve a standing review
Inventory your service principals and enterprise applications, remove consents you no longer use, and cut permissions to the minimum. Use application management policies to limit client secrets and their lifetimes, apply Conditional Access to workload identities where your licensing allows, and alert on any app that adds credentials to another app or assigns itself a role (Huntress, Sophos). Ask critical vendors how they protect the credentials their apps hold in your tenant and how quickly they will tell you if those credentials are abused.
Harden the cloud paths agents have already used
Require IMDSv2 on AWS instances, restrict who can read LAPS attributes, review Cognito self-registration and client-writable attributes, and rotate access keys every 90 days or retire them (SOCRadar, Datadog). If you run AI gateways or development platforms such as LiteLLM, Dify, Langflow, or Open WebUI, treat them as production systems. Remove default credentials, patch known flaws including CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow, restrict unauthenticated interfaces, and watch for unexpected cost or usage that suggests someone is spending your model credits (Security Affairs).
Restrict outbound traffic from internet-facing servers
Storm-3358's RAT staging depended on compromised servers being able to pull files from attacker infrastructure, so egress filtering on those hosts would have complicated it. Ask whether each internet-facing host needs outbound internet access at all.
Use deception against the agent's decision-making
An agent pursues an objective and accepts evidence that the objective is met. An experienced human operator who lands on a host that looks too valuable and too easy gets suspicious, and an agent may not. Microsoft's guidance leans on this: honeypots and decoy resources give the agent attractive targets with no value, tarpits consume its time and budget, and fake wins show convincing signs that its goal is complete, which can make it stop on a false success.
In Microsoft's testing, these controls reduced attack traffic reaching real assets and produced high-confidence signal useful for attribution. Confirm that decoys are believable, monitored, and connected to response, not just deployed. Automation can also be noisy: one outlet, summarizing Microsoft's separate credential-abuse research, reported ARTEX activity that generated 240 log events and errors that may have come from agents interfering with each other (Remio). That detail is not from the Storm-3358 report, but it is a reminder that speed and stealth are different things.
Test against the same tools
Unit 42 recommends running open-weight offensive frameworks internally to find exposures before external agents do. If you do, treat open-source pentest harnesses as hostile code: isolate them, keep production credentials out, enforce scope in configuration rather than in prompts, and review forks before running them, given Sophos's report of a trojanized ARTEX Docker image (Zscaler, Sophos). Register authorized agent infrastructure with your SOC so real activity can be told apart from testing. Purple team exercises are a practical way to confirm that the behavioral detections above fire. Unit 42 also suggests returning identical generic error codes or silently throttling, which leaves an agent without the feedback it uses to adapt.
Measure time to next step, not just findings
In an assessment, record the time from the first malicious request to the first alert, to triage, and to containment, and compare it with how long the attacker needs for the next step. If detection is measured in hours and the attacker's next step in minutes, that gap is the finding.
Test the entry points Storm-3358 used
Inventory every internet-facing file upload function and test whether it accepts files without authentication or executes what it stores. Test applications that fetch URLs on a user's behalf for SSRF, including which internal services they can reach. Then check what a compromised web server can do: whether a process spawning a shell raises an alert, and whether the server can reach arbitrary internet hosts with curl or wget.
Run purple team exercises with automated or agentic tooling to see whether rate- and behavior-based detections fire, rather than relying on IOC matches that an attacker will rotate.
Set rules of engagement before using agents on client environments
Offensive teams, ours included, are adopting agentic tools because they work. Using them requires hard scope enforcement, rate limits, a kill switch, and a human approval gate before any destructive or lateral step. An agent that wanders out of scope is still the tester's agent.
Plan for a shorter clock
If privileged access is possible within three hours, a runbook that starts the next business day will not help. Decide in advance who can disable a service principal or revoke a key at 2 a.m., automate that first step where you can with analyst confirmation afterward (Sophos), keep identity and Graph activity logs flowing to your SIEM, and rehearse the scenario in a tabletop exercise. Vapor Panda's operators told their agents to skip hosts running EDR, so confirm that sensor coverage includes servers you do not think of as internet-facing.
Questions to put to your team
None of these require a new budget line to answer. They show whether your current program is calibrated for an attacker who does not pause.
Do we have a current inventory of every internet-facing file upload and URL-fetch function, and when were they last tested?
If one of our web servers started downloading files from an unknown host, what would alert, who would see it, and how fast?
Can our internet-facing hosts reach the open internet outbound, and do they need to?
Are any credentials or secrets exposed in public code repositories, and how would we know?
Do we have any deception in place, and is it connected to response?
Echelon works across each of these areas. Defensive Security + Hardening assesses Microsoft 365, Entra, and cloud configurations and runs threat hunts for the behaviors above. Offensive Security + Adversary Simulation, including purple teaming, tests whether your detections fire against agent-style activity. Risk Advisory + GRC and the vCISO-Led Security Team as a Service cover third-party risk, AI governance, and incident response planning, and Managed Security Services provides 24/7 monitoring and response. Findings from offensive testing, such as exposed upload or SSRF paths, feed Defensive Security + Hardening and Managed Security Services detection tuning, and purple teaming confirms that the loop works. For financial institutions, Risk Advisory + GRC and the vCISO-Led Security Team as a Service map the gaps to FFIEC and GLBA obligations.
ARTEX makes familiar attacks faster, and the defenses against them are familiar too. This week, hunt your identity and cloud logs back to September 1 for the behaviors and indicators above, review every third-party application with permissions in your tenant, and confirm who can contain an account or service principal outside business hours. The indicators will age quickly. The behaviors and the fixes will last longer. Then test how long your detection takes compared with how long an attacker's next step takes.