Skip to navigation

Threat reportVulnerabilityTL-2026-2642

SalesBleed: Salesforce Agentforce vulnerabilities enable zero-click CRM data theft and trusted-agent Slack phishing

highACTIVE

Zenity Labs disclosed SalesBleed, a set of three vulnerabilities in Salesforce Agentforce that let an untrusted lead hijack trusted AI agents. Attackers can silently exfiltrate CRM data (zero-click) via Web-to-Lead indirect prompt injection and Trusted URLs bypasses, and impersonate the agent to phish employees from within trusted internal Slack channels via the 'Reply to a Slack Thread' action.

Key facts

IdentifiersTL-2026-2642
Also known asSalesBleed, ForcedLeak, Prompt Mines
SeverityHIGH
Affected productsSalesforce Agentforce; Salesforce Einstein AI / Einstein Generative AI
Weaknesses (CWE)CWE-74, CWE-20, CWE-918
MalwareGozi
MITRE ATT&CK / ATLAS techniquesAML.T0010 AI Supply Chain Compromise, AML.T0043 Craft Adversarial Data, AML.T0051 LLM Prompt Injection, T1048 Exfiltration Over Alternative Protocol, T1071 Application Layer Protocol, T1190 Exploit Public-Facing Application, T1204 User Execution, T1213 Data from Information Repositories, T1566 Phishing, T1567 Exfiltration Over Web Service, T1583 Acquire Infrastructure, T1584 Compromise Infrastructure and 1 more in the attack flow
MotivationUNKNOWN
Target sectorsfinance, health, retail, technology, telecoms, manufacturing
Target regionsGlobal
Detection rules9 (1 shown free)
Sources10 cited sources

Detection coverage

As of 2026-09-24, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2642 across Splunk SPL, Microsoft KQL and Sigma, covering 5 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: Salesforce Web-to-Lead Form with LLM Prompt-Injection Payload — Exploitation

Detects inbound Salesforce Web-to-Lead / Web-to-Case form submissions that carry indirect LLM prompt-injection payload markers (the SalesBleed/ForcedLeak planting step). An attacker with only public Web-to-Lead access embeds dormant malicious instructions in a Description/comment field; this detection surfaces the initial-access injection before it can hijack an Agentforce agent.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1190, T1204, T0051
  • Data sources: web_proxy, waf, salesforce_web_to_lead
  • Severity: high
  • Confidence: high

Splunk SPL

index=web sourcetype="access_combined" method=POST (uri_path="*WebToLead*" OR uri_path="*WebToCase*")
| rex field=_raw "(?i)(?<marker>###IMPORTANT INSTRUCTIONS|###INSTRUCTIONS|ignore (all )?previous instructions|disregard (all )?previous|you are now|system prompt|roleplay)"
| where isnotnull(marker)
| stats count by _time src_ip clientip uri_path marker
| table _time src_ip uri_path marker count

Microsoft KQL

// Web proxy / WAF access log (CommonSecurityLog) capturing Salesforce Web-to-Lead form POSTs.
CommonSecurityLog
| where TimeGenerated > ago(24h)
| where RequestMethod == "POST" and (RequestURL contains "WebToLead" or RequestURL contains "WebToCase")
| extend marker = iff(
    Message matches regex @"(?i)(###IMPORTANT INSTRUCTIONS|###INSTRUCTIONS|ignore (all )?previous instructions|disregard (all )?previous|you are now|system prompt|roleplay)",
    "prompt_injection_marker", "")
| where marker != ""
| project TimeGenerated, SourceIP, DestinationIP, RequestURL, marker, Message
| order by TimeGenerated desc

Sigma

title: Salesforce Web-to-Lead Form Submission with LLM Prompt-Injection Payload
id: 9f7c2b4a-1d3e-4a5f-8b6c-0a1b2c3d4e5f
status: experimental
description: >
  Detects Salesforce Web-to-Lead / Web-to-Case form submissions carrying indirect
  LLM prompt-injection payload markers (SalesBleed / ForcedLeak pattern). An attacker
  with only Web-to-Lead access plants dormant instructions that later hijack an
  Agentforce agent to exfiltrate CRM data or phish from a trusted identity.
references:
  - https://www.darkreading.com/application-security/salesbleed-exploits-salesforce-agents-slack-phishing
  - https://zenity.io/blog/security/zero-click-attacks-when-ttps-resurface-across-platforms
  - https://www.noma.security/blog/forcedleak-agent-risks-exposed-in-salesforce-agentforce
author: Threadlinqs Team
date: 2026/09/24
tags:
  - attack.initial_access
  - attack.execution
  - attack.t1190
  - attack.t1204
  - attack.aml.t0051
  - attack.aml.t0043
  - attack.aml.t0010
  - attack.ai.llm_prompt_injection
logsource:
  category: webserver
  product: generic
detection:
  selection_lead:
    cs-uri-query|contains:
      - WebToLead
      - WebToCase
  selection_payload:
    cs-uri-query|contains|all:
      - '###IMPORTANT INSTRUCTIONS'
  selection_payload2:
    cs-uri-query|contains:
      - '###INSTRUCTIONS'
      - 'ignore previous'
      - 'ignore all previous'
      - 'disregard'
      - 'system prompt'
      - 'you are now'
  condition: selection_lead and (selection_payload or selection_payload2)
falsepositives:
  - Legitimate marketing automation importing records whose descriptions contain instruction-like boilerplate
  - Security researchers testing prompt-injection protections against their own tenant
level: high

Known false positives:

  • Legitimate CRM imports containing instruction-like field text
  • Authorized security testing of injection protections

Tuning notes: Capture the full POST body in the proxy/access log (not just URL) so payload markers in Description/Comment fields are visible; restrict to public-facing Web-to-Lead endpoints; raise severity if the same source IP submits repeated injection-pattern records.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Salesforce Web-to-Lead Form with LLM Prompt-Injection Payload — Exploitation · unlock with TL Creds
  • Agent-Initiated Outbound Exfiltration to Unrecognized TLD / Expired Trusted Domain — Post-Exploitation · unlock with TL Creds
  • Slack Agent-Initiated Message / URL-Unfurling to Suspicious External Domain — C2 & Lateral Movement · unlock with TL Creds
  • Salesforce Web-to-Lead Form with LLM Prompt-Injection Payload — Exploitation · unlock with TL Creds
  • Agent-Initiated Outbound Exfiltration to Unrecognized TLD / Expired Trusted Domain — Post-Exploitation · unlock with TL Creds
  • Slack Agent-Initiated Message / URL-Unfurling to Suspicious External Domain — C2 & Lateral Movement · unlock with TL Creds
  • Agent-Initiated Outbound Exfiltration to Unrecognized TLD / Expired Trusted Domain — Post-Exploitation · unlock with TL Creds
  • Slack Agent-Initiated Message / URL-Unfurling to Suspicious External Domain — C2 & Lateral Movement · unlock with TL Creds

Indicators of compromise

As of 2026-09-24, Threadlinqs Intelligence tracks 5 indicators of compromise for TL-2026-2642.

Sample indicators (1 of 5)

  • Domain: my-salesforce-cms[.]com

Defanged, TLP:CLEAR. Full set of 5 indicators: unlock with TL Creds.

5 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Salesforce acknowledged the flaws (no CVE numbers assigned), found no evidence of real-world exploitation, and shipped fixes; all three were verified fixed by September 21, 2026.

SalesBleed is a set of three related security vulnerabilities in Salesforce Agentforce, an AI-agent platform that embeds into the Salesforce CRM (and Einstein AI), disclosed by Zenity Labs on September 24, 2026. The flaws share a root cause in how Agentforce governs and sanitizes external URLs and how it handles instructions carried in untrusted data. They allow an attacker with only the ability to submit a public Web-to-Lead form to plant a dormant indirect prompt injection inside the CRM, which later detonates when an employee legitimately asks an Agentforce agent about leads or records.

The first vulnerability is zero-click CRM data exfiltration through a Trusted URLs bypass. Salesforce's Trusted URLs control is meant to restrict which external destinations Agentforce agents may access and to redact links/images that point to untrusted URLs. Zenity found the regex-based redaction failed to register hostnames ending in unrecognized top-level domains, and that certain characters interfered with how URLs were parsed, letting crafted URLs evade redaction. Using this, an attacker's injected instructions cause the agent to query the Accounts table (via the subagent's Query Records tool), take stolen fields (company name, deal size, email addresses), embed them as a subdomain string for an attacker-controlled hostname, and print the result as an HTML <img src> tag. When the frontend renders it, the browser or agent infrastructure makes a DNS query / HTTP request to the attacker's server, silently exfiltrating the data with no click from the victim. A related second zero-click variant abuses Slack's URL unfurling: specially constructed links cause Slack to initiate requests that carry CRM data to attacker-controlled infrastructure as soon as the links appear in a channel.

The third vulnerability is trusted-agent impersonation for phishing. The Agentforce-Slack integration's 'Reply to a Slack Thread' action did not require user confirmation and did not surface attribution, so an agent (hijacked via indirect prompt injection, or driven by an insider) could post phishing messages under the trusted agent's identity from within internal Slack channels, with the sender remaining effectively anonymous. Salesforce subsequently updated default settings so certain Agentforce actions in Slack require user confirmation.

The findings build on earlier related research on the same platform. Noma Security's ForcedLeak (disclosed September 2025, CVSS-equivalent 9.4) demonstrated a nearly identical Web-to-Lead indirect prompt injection against Agentforce that exfiltrated CRM data by exploiting a Salesforce Content Security Policy that whitelisted *.my-salesforce-cms.com - a domain that had expired and become purchasable - turning it into a 'trusted' exfiltration channel. Salesforce responded by enforcing Trusted URL allowlisting for Agentforce and Einstein AI on September 8, 2025, the very control SalesBleed later found ways to bypass. Zenity also demonstrated related 'Prompt Mines' 0-click indirect prompt injections against Salesforce Einstein (shown at Black Hat 2025).

Salesforce acknowledged the flaws and responsibly shipped fixes: the Trusted URLs bypass was fixed (confirmed August 19, 2026) and all three flaws were confirmed fixed by September 21, 2026. No CVE numbers were assigned and no evidence of real-world exploitation had been observed at reporting time. The research community notes this class of AI-agent prompt-injection and URL-governance weakness is not Salesforce-specific and may recur across other AI agent platforms, so defenders should treat agents as production components, enforce least-privilege URL allowlists, monitor for anomalous outbound image/DNS requests, and require human confirmation for agent-initiated external actions.

MITRE ATT&CK / ATLAS techniques by tactic:

initial-access

AML.T0010 AI Supply Chain Compromise

AI Attack Staging

AML.T0043 Craft Adversarial Data

execution

AML.T0051 LLM Prompt Injection

Exfiltration

T1048 Exfiltration Over Alternative Protocol; T1567 Exfiltration Over Web Service

Command and Control

T1071 Application Layer Protocol

Initial Access

T1190 Exploit Public-Facing Application; T1566 Phishing

Execution

T1204 User Execution

Collection

T1213 Data from Information Repositories

Resource Development

T1583 Acquire Infrastructure; T1584 Compromise Infrastructure

Defense Evasion

T1684.001 Impersonation

Remediation

Patches

  • Apply Salesforce Trusted URLs Enforcement and the updated default settings for Agentforce Slack actions (all three SalesBleed flaws verified fixed by 2026-09-21)

Immediate actions

  • Enable and verify Trusted URL allowlists for Agentforce and Einstein AI, adding all external domains agents legitimately call
  • Review and tighten Content-Security-Policy directives (especially img-src and connect-src) for agent-generated content
  • Audit existing lead/case/account data for suspicious Web-to-Lead and Web-to-Case submissions carrying injected instructions
  • Enable user confirmation for Agentforce-to-Slack actions such as 'Reply to a Slack Thread'

Workarounds

  • Restrict Web-to-Lead field lengths to reduce injection payload capacity
  • Disable or restrict unused agent-to-Slack and image-generation actions
  • Monitor for outbound requests to unrecognized TLDs / suspicious hostnames from Salesforce and Slack infrastructure

Longer-term hardening

  • Deploy prompt-injection detection and input validation on user-controlled CRM fields (Web-to-Lead, Email-to-Case, Web-to-Case)
  • Treat AI agents as production components: maintain an AI agent inventory/BOM and threat-model agent tool-calling
  • Add runtime tool-calling and exfiltration guardrails (anomalous outbound DNS/image/HTTP from agent infrastructure)
  • Apply least-privilege data access to agents so they cannot query records outside their required scope

Affected and fixed versions

  • Salesforce — Agentforce
    Vulnerable versions: pre-Trusted-URLs-enforcement and pre-Sept-2026-patch builds
    Fixed in: Trusted URLs Enforcement (2025-09-08) and SalesBleed fixes confirmed 2026-09-21
  • Salesforce — Einstein AI / Einstein Generative AI
    Vulnerable versions: builds before Trusted URL allowlisting enforcement
    Fixed in: Trusted URL allowlisting enforcement (2025-09-08)

Timeline

  • Noma Security reports the ForcedLeak indirect prompt injection chain in Salesforce Agentforce (related prior disclosure on the same platform).
  • Salesforce enforces Trusted URL allowlisting for Agentforce and Einstein Generative AI in response to ForcedLeak - the control SalesBleed later bypasses.
  • Zenity Labs responsibly discloses SalesBleed (three Salesforce Agentforce flaws) to Salesforce.
  • Salesforce confirms the Trusted URLs bypass vulnerability is fixed.
  • Salesforce confirms all three SalesBleed flaws are fixed; no evidence of real-world exploitation reported.
  • TL-Intel harness creates this threat record for defender awareness and Slack/Agentforce phishing detection.
  • Zenity Labs publicly discloses SalesBleed; Dark Reading and The Register report; Salesforce updates default settings so certain Agentforce actions in Slack require user confirmation.

Sources cited

Previous and next in vulnerability

Drafted by the Threadlinqs research pipeline from 10 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats