# [CRIT] Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE

**Source:** Unit 42 (Palo Alto)
**Published:** 2026-06-16
**Article:** https://unit42.paloaltonetworks.com/hijacking-vertex-ai-model/

## Threat Profile

Threat Research Center 
Threat Research 
Cloud Cybersecurity Research 
Cloud Cybersecurity Research 
Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE 
11 min read 
Related Products Cortex Cortex Cloud Unit 42 AI Security Assessment Unit 42 Incident Response 
By: Ori Hadad 
Published: June 16, 2026 
Categories: Cloud Cybersecurity Research 
Threat Research 
Tags: Bucket squatting 
Google Cloud 
Joblib 
Python 
RCE 
SDKs 
Vertex AI 
Vulnerability 
Executive Summary 
We…

## Indicators of Compromise (high-fidelity only)

- **CVE:** `CVE-2026-2473`

## MITRE ATT&CK Techniques

- **T1190** — Exploit Public-Facing Application
- **T1566.002** — Spearphishing Link
- **T1204.001** — User Execution: Malicious Link
- **T1059.001** — PowerShell
- **T1566.001** — Spearphishing Attachment
- **T1204.002** — User Execution: Malicious File
- **T1059.005** — Visual Basic
- **T1218** — System Binary Proxy Execution
- **T1021.002** — SMB/Windows Admin Shares
- **T1569.002** — Service Execution
- **T1528** — Steal Application Access Token
- **T1098.001** — Account Manipulation: Additional Cloud Credentials
- **T1219** — Remote Access Software
- **T1195.002** — Compromise Software Supply Chain
- **T1059.006** — Command and Scripting Interpreter: Python
- **T1078.004** — Valid Accounts: Cloud Accounts
- **T1525** — Implant Internal Image
- **T1530** — Data from Cloud Storage

## Kill chain phases observed

_(none detected from narrative keywords)_

## Recommended hunts

### Vulnerable google-cloud-aiplatform SDK install (CVE-2026-2473 — pre-1.148.0)

`UC_42_8` · phase: **weapon** · confidence: **High** · AI-generated for this article

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count min(_time) as firstTime max(_time) as lastTime from datamodel=Endpoint.Processes where Processes.process_name IN ("pip.exe","pip3.exe","python.exe","python3.exe","pip","pip3","python","python3") Processes.process="*google-cloud-aiplatform*" by Processes.dest Processes.user Processes.process Processes.parent_process_name
| `drop_dm_object_name(Processes)`
| rex field=process "google-cloud-aiplatform==(?<sdk_version>1\.(?:13[9]|14[0-7])\.\d+)"
| where isnotnull(sdk_version)
| `security_content_ctime(firstTime)`
| `security_content_ctime(lastTime)`
```

**Defender KQL:**
```kql
DeviceProcessEvents
| where Timestamp > ago(14d)
| where FileName in~ ("pip.exe","pip3.exe","python.exe","python3.exe","pip","pip3","python","python3")
| where ProcessCommandLine has "google-cloud-aiplatform"
| where ProcessCommandLine matches regex @"google-cloud-aiplatform==1\.(139|140|141|142|143|144|145|146|147)\."
| extend VulnerableVersion = extract(@"google-cloud-aiplatform==(1\.\d+\.\d+)", 1, ProcessCommandLine)
| project Timestamp, DeviceName, AccountName, AccountUpn, VulnerableVersion, ProcessCommandLine, FolderPath, InitiatingProcessFileName, InitiatingProcessCommandLine
| order by Timestamp desc
```

### Vertex AI default staging bucket squatting — predictable bucket pattern access

`UC_42_9` · phase: **delivery** · confidence: **Medium** · AI-generated for this article

**Splunk SPL (CIM):**
```spl
`gcp_audit_logs` data.protoPayload.serviceName="storage.googleapis.com"
| rex field=data.protoPayload.resourceName "projects/_/buckets/(?<bucket_name>[^/]+)"
| where match(bucket_name, "-vertex-staging-(us|europe|asia|australia|northamerica|southamerica)")
| eval principal=data.protoPayload.authenticationInfo.principalEmail
| where NOT match(principal, "gcp-sa-aiplatform\.iam\.gserviceaccount\.com$")
| stats min(_time) as firstSeen max(_time) as lastSeen values(data.protoPayload.methodName) as methods values(principal) as principals count by bucket_name
| `security_content_ctime(firstSeen)`
| `security_content_ctime(lastSeen)`
```

### Phishing-link click correlated to endpoint execution

`UC_PHISH_LINK` · phase: **delivery** · confidence: **High**

**Splunk SPL (CIM):**
```spl
``` Phishing-link click that drives endpoint execution within 60s ```
| tstats `summariesonly` earliest(_time) AS click_time
    from datamodel=Web
    where Web.action="allowed"
    by Web.src, Web.user, Web.dest, Web.url
| `drop_dm_object_name(Web)`
| rename user AS recipient, dest AS clicked_domain, url AS clicked_url
| join type=inner recipient
    [| tstats `summariesonly` count
         from datamodel=Email.All_Email
         where All_Email.action="delivered" AND All_Email.url!="-"
         by All_Email.recipient, All_Email.src_user, All_Email.url, All_Email.subject
     | `drop_dm_object_name(All_Email)`
     | rex field=url "https?://(?<email_domain>[^/]+)"
     | rename recipient AS recipient]
| join type=inner src
    [| tstats `summariesonly` earliest(_time) AS exec_time
         values(Processes.process) AS exec_cmd, values(Processes.process_name) AS exec_proc
         from datamodel=Endpoint.Processes
         where Processes.parent_process_name IN ("chrome.exe","msedge.exe","firefox.exe",
                                                   "outlook.exe","brave.exe","arc.exe")
           AND Processes.process_name IN ("powershell.exe","pwsh.exe","cmd.exe","mshta.exe",
                                            "rundll32.exe","regsvr32.exe","wscript.exe",
                                            "cscript.exe","bitsadmin.exe","certutil.exe",
                                            "curl.exe","wget.exe")
         by Processes.dest, Processes.user
     | `drop_dm_object_name(Processes)`
     | rename dest AS src]
| eval delta_sec = exec_time - click_time
| where delta_sec >= 0 AND delta_sec <= 60
| table click_time, exec_time, delta_sec, recipient, src, src_user, subject,
        clicked_domain, clicked_url, exec_proc, exec_cmd
| sort - click_time
```

**Defender KQL:**
```kql
// Phishing-link click that drives endpoint execution within 60s.
// Far higher fidelity than "every clicked URL" — most legitimate clicks
// never spawn a non-browser child process, so the join eliminates the
// 99% of noise that makes a raw click query unactionable.
let LookbackDays = 7d;
let SuspectClicks = UrlClickEvents
    | where Timestamp > ago(LookbackDays)
    | where AccountName !endswith "$"
    | where ActionType in ("ClickAllowed","ClickedThrough")
    | join kind=inner (
        EmailEvents
        | where Timestamp > ago(LookbackDays)
        | where DeliveryAction == "Delivered"
        | where EmailDirection == "Inbound"
        | project NetworkMessageId, Subject, SenderFromAddress, SenderFromDomain,
                  RecipientEmailAddress, EmailTimestamp = Timestamp
      ) on NetworkMessageId
    | join kind=leftouter (
        EmailUrlInfo | project NetworkMessageId, Url, UrlDomain
      ) on NetworkMessageId, Url
    | project ClickTime = Timestamp, AccountUpn, IPAddress, Url, UrlDomain,
              Subject, SenderFromAddress, SenderFromDomain, RecipientEmailAddress,
              ActionType;
// Correlate to a non-browser child process spawned within 60 seconds on
// the recipient's device.
DeviceProcessEvents
| where Timestamp > ago(LookbackDays)
| where InitiatingProcessFileName in~ ("chrome.exe","msedge.exe","firefox.exe",
                                         "outlook.exe","brave.exe","arc.exe")
| where FileName in~ ("powershell.exe","pwsh.exe","cmd.exe","mshta.exe",
                        "rundll32.exe","regsvr32.exe","wscript.exe","cscript.exe",
                        "bitsadmin.exe","certutil.exe","curl.exe","wget.exe")
| join kind=inner SuspectClicks on $left.AccountName == $right.AccountUpn
| where Timestamp between (ClickTime .. ClickTime + 60s)
| project ClickTime, ProcessTime = Timestamp,
          DelaySec = datetime_diff('second', Timestamp, ClickTime),
          DeviceName, AccountName, RecipientEmailAddress, SenderFromAddress,
          Subject, Url, UrlDomain, ActionType,
          FileName, ProcessCommandLine, InitiatingProcessFileName
| order by ClickTime desc
```

### Email attachment opened from external sender

`UC_PHISH_ATTACH` · phase: **delivery** · confidence: **High**

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count
    from datamodel=Email.All_Email
    where All_Email.file_name!="-"
    by All_Email.src_user, All_Email.recipient, All_Email.file_name, All_Email.subject
| rename All_Email.recipient as user
| join type=inner user
    [| tstats `summariesonly` count
        from datamodel=Endpoint.Processes
        where Processes.parent_process_name IN ("OUTLOOK.EXE","winword.exe","excel.exe","powerpnt.exe")
          AND Processes.process_name IN ("cmd.exe","powershell.exe","wscript.exe","cscript.exe","mshta.exe","rundll32.exe","regsvr32.exe")
        by Processes.dest, Processes.user, Processes.parent_process_name, Processes.process_name, Processes.process
     | rename Processes.user as user]
```

**Defender KQL:**
```kql
let LookbackDays = 7d;
let MalAttachments = EmailAttachmentInfo
    | where Timestamp > ago(LookbackDays)
    | where AccountName !endswith "$"
    | project NetworkMessageId, RecipientEmailAddress,
              AttachmentFileName = FileName, AttachmentSHA256 = SHA256;
DeviceProcessEvents
| where Timestamp > ago(LookbackDays)
| where InitiatingProcessFileName in~ ("OUTLOOK.EXE","winword.exe","excel.exe","powerpnt.exe")
| where FileName in~ ("cmd.exe","powershell.exe","wscript.exe","cscript.exe",
                      "mshta.exe","rundll32.exe","regsvr32.exe")
| join kind=inner MalAttachments on $left.AccountUpn == $right.RecipientEmailAddress
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine,
          InitiatingProcessFileName, AttachmentFileName, AttachmentSHA256
```

### Office app spawning script/LOLBin child process

`UC_OFFICE_CHILD` · phase: **exploit** · confidence: **High**

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count min(_time) as firstTime max(_time) as lastTime
    from datamodel=Endpoint.Processes
    where Processes.parent_process_name IN ("winword.exe","excel.exe","powerpnt.exe","outlook.exe","onenote.exe","mspub.exe","visio.exe")
      AND Processes.process_name IN ("cmd.exe","powershell.exe","pwsh.exe","wscript.exe","cscript.exe","mshta.exe","rundll32.exe","regsvr32.exe","wmic.exe","bitsadmin.exe","certutil.exe")
    by Processes.dest, Processes.user, Processes.parent_process_name, Processes.process_name, Processes.process
| `drop_dm_object_name(Processes)`
| `security_content_ctime(firstTime)` | `security_content_ctime(lastTime)`
```

**Defender KQL:**
```kql
DeviceProcessEvents
| where Timestamp > ago(7d)
| where AccountName !endswith "$"
| where InitiatingProcessFileName in~ ("winword.exe","excel.exe","powerpnt.exe","outlook.exe","onenote.exe","mspub.exe","visio.exe")
| where FileName in~ ("cmd.exe","powershell.exe","pwsh.exe","wscript.exe","cscript.exe","mshta.exe","rundll32.exe","regsvr32.exe","wmic.exe","bitsadmin.exe","certutil.exe")
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, FileName, ProcessCommandLine
```

### Remote service execution — PsExec / SMB lateral movement

`UC_LATERAL_PSEXEC` · phase: **actions** · confidence: **High**

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count min(_time) as firstTime max(_time) as lastTime
    from datamodel=Endpoint.Processes
    where Processes.process_name IN ("psexec.exe","psexesvc.exe","paexec.exe","smbexec.py")
       OR (Processes.process_name="wmic.exe" AND Processes.process="*/node:*")
    by Processes.dest, Processes.user, Processes.process_name, Processes.process, Processes.parent_process_name
| `drop_dm_object_name(Processes)`
```

**Defender KQL:**
```kql
DeviceProcessEvents
| where Timestamp > ago(7d)
| where AccountName !endswith "$"
| where FileName in~ ("psexec.exe","psexesvc.exe","paexec.exe","smbexec.py")
   or (FileName =~ "wmic.exe" and ProcessCommandLine has "/node:")
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessFileName
| order by Timestamp desc
```

### OAuth consent / suspicious app grant

`UC_OAUTH_ABUSE` · phase: **actions** · confidence: **High**

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count min(_time) as firstTime max(_time) as lastTime
    from datamodel=Authentication.Authentication
    where Authentication.action="success"
      AND Authentication.signature IN (
        "Consent to application",
        "Add app role assignment grant to user",
        "Add OAuth2PermissionGrant",
        "Add delegated permission grant")
    by Authentication.user, Authentication.app, Authentication.src, Authentication.signature
| `drop_dm_object_name(Authentication)`
```

**Defender KQL:**
```kql
CloudAppEvents
| where Timestamp > ago(7d)
| where ActionType in ("Consent to application.","Add OAuth2PermissionGrant.","Add delegated permission grant.")
| project Timestamp, AccountObjectId, AccountDisplayName, ActivityType,
          ActivityObjects, IPAddress, UserAgent
```

### RMM tool installed by non-IT user — remote-access utility for hands-on-keyboard

`UC_RMM_TOOLS` · phase: **install** · confidence: **High**

**Splunk SPL (CIM):**
```spl
| tstats `summariesonly` count min(_time) as firstTime max(_time) as lastTime
    from datamodel=Endpoint.Processes
    where Processes.process_name IN ("AnyDesk.exe","TeamViewer.exe","TeamViewer_Service.exe",
        "ScreenConnect.ClientService.exe","ConnectWiseControl.ClientService.exe",
        "atera_agent.exe","SplashtopStreamer.exe","RustDesk.exe","NinjaOne.exe","kaseya*.exe")
    by Processes.dest, Processes.user, Processes.process_name, Processes.process, Processes.parent_process_name
| `drop_dm_object_name(Processes)`
```

**Defender KQL:**
```kql
DeviceProcessEvents
| where Timestamp > ago(7d)
| where AccountName !endswith "$"
| where FileName in~ ("AnyDesk.exe","TeamViewer.exe","TeamViewer_Service.exe",
        "ScreenConnect.ClientService.exe","ConnectWiseControl.ClientService.exe",
        "atera_agent.exe","SplashtopStreamer.exe","RustDesk.exe","NinjaOne.exe")
   or FileName matches regex @"(?i)kaseya.*\.exe"
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine
```

### Article-specific behavioural hunt — Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE

`UC_42_7` · phase: **exploit** · confidence: **High**

**Splunk SPL (CIM):**
```spl
``` Article-specific bespoke detection — Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE ```
| tstats `summariesonly` count earliest(_time) AS firstTime latest(_time) AS lastTime
    from datamodel=Endpoint.Processes
    where (Processes.process_name IN ("gcs_utils.py"))
    by Processes.dest, Processes.user, Processes.process_name,
       Processes.process, Processes.parent_process_name, Processes.process_path
| `drop_dm_object_name(Processes)`
| `security_content_ctime(firstTime)`
| append [
| tstats `summariesonly` count
    from datamodel=Endpoint.Filesystem
    where Filesystem.action IN ("created","modified")
      AND (Filesystem.file_name IN ("gcs_utils.py"))
    by Filesystem.dest, Filesystem.user, Filesystem.process_name,
       Filesystem.file_path, Filesystem.file_name
| `drop_dm_object_name(Filesystem)`
]
```

**Defender KQL:**
```kql
// Article-specific bespoke detection — Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE
// Hunts the actual binaries / paths / commandline fragments named
// in the article instead of a generic technique-class template.
DeviceProcessEvents
| where Timestamp > ago(30d)
| where (FileName in~ ("gcs_utils.py"))
| project Timestamp, DeviceName, AccountName, FileName,
          FolderPath, ProcessCommandLine,
          InitiatingProcessFileName, InitiatingProcessCommandLine
| order by Timestamp desc

// File-creation events for the named binaries / paths
DeviceFileEvents
| where Timestamp > ago(30d)
| where ActionType in ("FileCreated","FileModified")
| where (FileName in~ ("gcs_utils.py"))
| project Timestamp, DeviceName, AccountName, FolderPath,
          FileName, ActionType, InitiatingProcessFileName,
          InitiatingProcessCommandLine
| order by Timestamp desc
```

### IOC-driven hunts (use shared templates)

These are standard IOC-substitution hunts — the canonical SPL and KQL live once in [`_TEMPLATES.md`](../_TEMPLATES.md), so we don't repeat the same boilerplate on every CVE / hash / network-IOC briefing.

- **Asset exposure — vulnerability matches article CVE(s)** ([template](../_TEMPLATES.md#asset-exposure)) — phase: **recon**, confidence: **High**
  - CVE(s): `CVE-2026-2473`


## Why this matters

Severity classified as **CRIT** based on: CVE present, 10 use case(s) fired, 18 technique(s) inferred. Read the full article for actor attribution, tooling details, and any defanged IOCs in the body that aren't visible in the RSS summary.