China Hackers Used DeepSeek AI to Hit 200+ Organisations: CrowdStrike
CrowdStrike says China-linked hackers used ARTEX, an AI agent framework running mainly on DeepSeek, to target more than 200 organisations in 11 days. Here is what it found and a practical defence checklist for Indian businesses.
Image: TeQBlogs
CrowdStrike says a China-linked hacking group it calls VAPOR PANDA used an AI agent framework named ARTEX to target more than 200 organisations in at least 14 countries in just 11 days. The framework ran mainly on DeepSeek, the Chinese-built AI model, plus a second commercial AI model that CrowdStrike could not identify. A single human operator steered many intrusions at once. For Indian businesses, the message is plain: attacks that once needed a large team can now run at machine speed, so basic security gaps get found and used faster than ever.
Quick summary
| Detail | What CrowdStrike reported |
|---|---|
| Hacking group | VAPOR PANDA, a China-nexus espionage group active since 2021 |
| AI tool | ARTEX, a multi-agent AI framework |
| AI models | DeepSeek (main model) and an unconfirmed subscription-based commercial model |
| Scale | 200+ organisations, at least 14 countries |
| Time frame | 11 days, starting 17 October 2026 |
| Confirmed break-ins | 9 organisations, plus password-spraying against at least 12 more |
| India mentioned? | No. India is not named in the report. |
What CrowdStrike found
In a threat-intelligence blog dated 29 October 2026, CrowdStrike described this as the first time it has seen VAPOR PANDA use AI-driven intrusion tooling. The group normally focuses on building long-term, quiet access inside its targets.
ARTEX works like a small team of AI workers. A main agent hands tasks to “planner” and “worker” agents. They all share a common map of facts, findings and goals, and they move step by step from first access to reconnaissance and then deeper into the network. A human operator, writing instructions in Chinese, stepped in at key decision points. CrowdStrike says the planning agents in a single intrusion consumed about 1.4 billion input tokens. That is a measure of how much the AI “read” and reasoned over.
Also read: Binance Hack 2026: What the $6.3 Billion Theft Means for Indian Users
In the nine confirmed break-ins, the attackers installed a backdoor called BLOODALCHEMY. It was first publicly documented in 2023, and the attackers hid it behind a renamed, legitimate program disguised as a Windows system file. CrowdStrike’s summary sums it up: “This campaign marks an escalation in China-nexus use of AI agents, from single intrusions to multiple simultaneous engagements directed by a human operator.”
ARTEX is not new. On 21 October 2026, Microsoft described ARTEX as an open-source penetration-testing framework that can use any AI model. Microsoft linked it to a different China-state-aligned actor, Storm-0940, and to a supply-chain compromise involving ConnectWise. CrowdStrike says its VAPOR PANDA campaign is separate from that activity. So at least two China-linked groups are now using the same AI attack tool.
Who was targeted
- Governments: foreign ministries in Latin America (two extensively compromised, a third likely), an interior ministry in the Middle East, and a European country’s foreign ministry, presidency, prime minister’s office and defence, justice, finance and intelligence bodies.
- Industry in the United States: defence, nuclear, semiconductor, pharmaceutical and financial organisations.
- Telecom providers: in the US, Belgium and the Philippines.
- Others: think tanks and a Greek government body.
CrowdStrike does not name India. But a tool that can scan and attack 200 organisations in under two weeks can easily be pointed at new countries and sectors.
How AI-powered attacks differ
| Traditional attack | AI-agent attack (as described by CrowdStrike) |
|---|---|
| A team of skilled hackers works on one target at a time | One operator directs many intrusions at the same time |
| Tries one path, and moves on slowly if it fails | Tries several paths in parallel (web apps, cloud, containers, Active Directory, email, VPN) and switches quickly |
| Manual reconnaissance takes days or weeks | Agents map the network and share findings automatically |
| Hiding from security tools is done by hand | Evasion is built into the workflow |
The last row matters most. According to CrowdStrike, the operator told the agents to find Windows machines that were not running its Falcon security sensor, and to focus further activity on those unprotected computers. The agents also exploited a flaw where SAML login messages were not properly signed, and abused cloud “metadata” services to grab credentials. They used old keys and exports that had never been changed, too. None of these are futuristic tricks. They are ordinary weaknesses, found and chained together much faster.
Why it matters for Indian businesses
Indian companies increasingly run on cloud platforms, Microsoft 365, VPNs and managed IT providers, which are exactly the surfaces ARTEX went after. Small and medium businesses often have one unpatched server, one shared admin password or one laptop without endpoint protection. An AI agent is very good at finding those. The ConnectWise case Microsoft described also shows that attackers can come in through a trusted IT vendor’s tools.
There is a legal angle too. Under CERT-In’s April 2022 directions, service providers, intermediaries, data centres, companies and government organisations must report listed cyber incidents to CERT-In within 6 hours of noticing them. They must also keep system logs for a rolling 180 days, stored in India. If you get hit, you will need those logs, and you will need to act fast.
Defence checklist for small and medium businesses
- Protect every device: Put endpoint detection and response (EDR) on every laptop and server, including internal servers that are not exposed to the internet. Attackers are now hunting for the ones you missed.
- Patch internet-facing systems first: VPNs, firewalls, mail servers and web apps. CrowdStrike specifically flags long-lived Linux servers.
- Use strong MFA everywhere: Prefer app-based or hardware-key MFA for email, VPN and cloud admin accounts. This blunts password spraying.
- Rotate old keys and passwords: After any suspected exposure, change signing keys, service account passwords and API keys. Remember the copies in old backups and directory exports.
- Check your login setup: Make sure single sign-on (SAML) signatures are always validated. On AWS, require IMDSv2 for instance metadata.
- Limit vendor and app permissions: Review what your IT provider’s tools and third-party apps can access in Microsoft 365 or Google Workspace, and remove what they don’t need.
- Watch for warning signs: Bursts of failed logins, new admin accounts, new app credentials and unusual mailbox access.
- Keep logs and an incident plan: Keep 180 days of logs, and know who will email incident@cert-in.org.in within 6 hours if something happens.
Frequently asked questions
Did DeepSeek itself hack these organisations?
No. DeepSeek is an AI model. The attackers used it as the “brain” inside the ARTEX framework, and a human operator directed the campaign.
Were any Indian organisations hit?
CrowdStrike’s report does not name India or any Indian organisation. It names the US, Belgium, the Philippines and Greece, plus several regions.
Is it safe for my company to use DeepSeek?
This report is about attackers misusing AI, not about risks to DeepSeek’s users. Follow your own data policy before putting company data into any AI tool.
Can AI also help defenders?
Yes. Security vendors use AI to spot unusual behaviour faster. But the basics still stop most attacks: patching, MFA, EDR and rotating credentials.
Bottom line
CrowdStrike’s findings show that state-backed hackers now use AI agents to run many attacks at once, and it expects China-linked groups to keep adopting tools like ARTEX over the next 12 months. The AI is new, but the weaknesses it exploits are old: missing security software, unpatched servers, weak logins and keys that never get changed. Fix those this month, and an AI attacker has far less to work with.
Sources
- CrowdStrike: VAPOR PANDA uses ARTEX to scale global campaign
- Microsoft Security Blog: Storm-0940 uses ARTEX in ConnectWise supply chain compromise
- CERT-In: Directions under Section 70B (28 April 2022)
About The Author
