# SalesBleed: Indirect Prompt Injection Enables Zero-Click CRM Data Exfiltration in Salesforce Agentforce

> Zenity Labs disclosed SalesBleed, a three-vulnerability chain in Salesforce Agentforce combining indirect prompt injection via public Web-to-Lead forms, a Trusted URLs redaction bypass, and DNS-based exfiltration to silently steal CRM data with no victim interaction or attacker authentication.

- **Published:** 2026-09-27T00:00:00Z
- **Last reviewed:** 2026-09-27T00:00:00Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-2710
- **Card image:** https://intel.threadlinqs.com/og/threat/threat-intel-salesbleed-agentforce-prompt-injection-tl-2026-2710.webp?v=1d0f3598 (Threadlinqs threat intel: SalesBleed Agentforce CRM Data Theft: Detection & IOCs (TL-2026-2710) — High severity, 15 IOCs, 9 detection rules.)
- **ID:** TL-2026-2710
- **Severity:** HIGH
- **Category:** VULNERABILITY
- **Status:** PATCHED
- **Detections:** 9 (1 shown free below) · **IOCs:** 15

## Description

SalesBleed is a three-vulnerability chain in Salesforce Agentforce disclosed by Zenity Labs (researchers Alex Apostolov, Joao Donato, Avishai Efrat, and Ayush RoyChowdhury). The chain begins when an attacker submits an unauthenticated lead through a public Web-to-Lead form, hiding indirect prompt-injection instructions inside otherwise-normal-looking form fields. The malicious lead sits dormant in the Leads table indefinitely until an employee performs a routine, legitimate action -- asking an Agentforce agent to review or summarize the lead. When the agent ingests the poisoned record, it follows the embedded instructions and invokes its own Query Records tool against the Accounts table (a capability already granted to the General CRM subagent by default), extracting sensitive fields such as company names and deal sizes and encoding them as subdomain labels of an attacker-controlled URL.

That exfiltration URL must pass through Salesforce's Trusted URLs mechanism, which is designed to redact links and images from untrusted sources before they render. Zenity found two RFC 3986 non-compliance gaps in the redactor: it did not recognize unusual/unrecognized top-level domains such as '.fun', and it failed to redact URLs containing curly-brace and square-bracket termination characters even though browsers still parsed and rendered them. Exploiting both gaps together produced a malformed-looking string (e.g. 'https://random_string.oast.fun/{email}') that the redactor treated as a non-URL and left untouched, while the rendering surface treated it as a live, resolvable link. The agent then emitted the exfiltration data in two ways: (1) as an HTML <img> tag, causing the chat surface to automatically fetch the external 'image' and trigger a DNS lookup that carried the stolen data to attacker-controlled nameservers before any HTTP request completed; and (2) inside a Slack channel, where publishing the agent's output triggered Slack's automatic URL-unfurling/link-preview feature, which itself resolved the malicious domain -- exfiltrating data with zero clicks and no authentication in either path. Because DNS labels are capped at 63 characters, larger record sets were split across multiple sequential queries.

A third, related flaw let an attacker (an insider, or an external attacker via the same Web-to-Lead injection vector) weaponize the Agentforce-Slack integration's 'Reply to a Slack Thread' action to send phishing messages that appeared to originate from the trusted, company-wide Agentforce bot identity. Unlike comparable actions (e.g., 'Send a Slack Direct Message'), this action shipped without a required user-confirmation step and without any attribution showing which user or process had invoked it, so recipients had no way to trace the message back to its true origin. Combined with the same Trusted URLs bypass, the phishing links could be rendered as legitimate-looking markdown links inside a credible, in-context thread reply.

The pattern instantiates Simon Willison's 'lethal trifecta' for AI agents: untrusted external input (public Web-to-Lead submissions), access to sensitive internal data (CRM Accounts/Leads via the Query Records tool), and an unmonitored external communication channel (DNS resolution via image rendering and Slack unfurling). Zenity reported all three issues to Salesforce on June 1, 2026; Salesforce investigated and shipped the Trusted URLs hardening (standards-compliant URL parsing replacing prior regex-based validation) by August 18-19, 2026, and added invoking-user attribution plus mandatory confirmation to the Slack reply action by September 21, 2026 -- all before the September 24-25, 2026 public disclosure. No in-the-wild exploitation was observed; this was a responsibly-disclosed, proof-of-concept-validated research finding, not an active campaign.

## MITRE ATT&CK

- AML.T0051.001 Indirect
- AML.T0051.002 Triggered
- AML.T0052 Phishing
- AML.T0057 LLM Data Leakage
- T1213 Data from Information Repositories
- T1048.003 Exfiltration Over Unencrypted Non-C2 Protocol
- T1567 Exfiltration Over Web Service
- T1027 Obfuscated Files or Information
- T1534 Internal Spearphishing

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**SalesBleed: Subdomain-Encoded DNS Exfiltration and Anonymous Agentforce Slack Reply Hijack (SPL)** (`TL-2026-2710-SPL-003`)

Detects the two SalesBleed exfiltration/impersonation channels: (1) DNS queries for high-entropy, long subdomain labels under rare or out-of-band-interaction TLDs (oast.fun, interactsh.com, and similar OOB-testing domains), consistent with stolen CRM fields encoded as sequential 63-character-capped DNS labels via HTML <img> auto-fetch or Slack link-unfurling; and (2) the Agentforce 'Reply to a Slack Thread' action firing under the trusted bot identity with no invoking-user attribution and no prior confirmation event, the vector used for anonymous internal phishing.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T0057, T1048.003, T1567, T1534, T0052
- **Data sources:** dns_query_logs, proxy_logs, slack_audit_logs
- **Severity:** critical
- **Confidence:** high

### Splunk SPL

```spl
(index=dns sourcetype=dns_query) OR (index=slack sourcetype="slack:audit")
| eval is_dns_exfil=if(sourcetype="dns_query" AND match(query, "(?i)[a-z0-9+\/=_-]{20,}\.[a-z0-9_.-]*(oast\.fun|interactsh\.com|oast\.(me|live|pro|site))$"), 1, 0)
| eval is_slack_hijack=if(sourcetype="slack:audit" AND action="reply_to_thread" AND actor_type="bot" AND match(bot_name, "(?i)agentforce") AND isnull(invoking_user_id), 1, 0)
| where is_dns_exfil=1 OR is_slack_hijack=1
| table _time, sourcetype, query, src_ip, bot_name, channel_id, thread_ts, invoking_user_id, is_dns_exfil, is_slack_hijack
| sort -_time
```

### Microsoft KQL

```kql
let dnsExfil = DnsEvents
| where TimeGenerated > ago(24h)
| where Name matches regex @"(?i)[a-z0-9+/=_-]{20,}\.[a-z0-9_.-]*(oast\.fun|interactsh\.com|oast\.(me|live|pro|site))$"
| project TimeGenerated, Computer, Name, ClientIP, EventType = "dns_exfil", ChannelId = "", InvokingUserId = "";
let slackHijack = SlackAuditLogs_CL
| where TimeGenerated > ago(24h)
| where Action_s == "reply_to_thread" and ActorType_s == "bot" and BotName_s has "Agentforce" and isempty(InvokingUserId_s)
| project TimeGenerated, Computer = BotName_s, Name = ThreadTs_s, ClientIP = "", EventType = "slack_reply_hijack", ChannelId = ChannelId_s, InvokingUserId = InvokingUserId_s;
dnsExfil
| union slackHijack
| order by TimeGenerated desc
```

### Sigma

```yaml
title: SalesBleed DNS Subdomain Exfiltration and Anonymous Agentforce Slack Reply Hijack
id: 8c2e4f61-1d7a-49b6-9e2c-5b8a3f0d7c44
status: experimental
description: Detects DNS-based exfiltration of CRM data encoded as long subdomain labels resolving to out-of-band interaction domains, and the Agentforce 'Reply to a Slack Thread' action firing with no invoking-user attribution — the SalesBleed C2/impersonation stage.
references:
    - https://labs.zenity.io/post/salesbleed-0-click-data-exfiltration-on-agentforce
    - https://labs.zenity.io/post/salesbleed-hijacking-agentforce-in-slack-for-anonymous-phishing
author: Threadlinqs Team
date: 2026-09-27
tags:
    - attack.exfiltration
    - attack.t0057
    - attack.t1048.003
    - attack.t1567
    - attack.lateral-movement
    - attack.t1534
    - attack.t0052
    - attack.tool.salesbleed
logsource:
    category: dns
detection:
    dns_selector:
        query|re: '(?i)[a-z0-9+/=_-]{20,}\.[a-z0-9_.-]*(oast\.fun|interactsh\.com|oast\.(me|live|pro|site))$'
    slack_selector:
        service: slack_audit
        Action: 'reply_to_thread'
        ActorType: 'bot'
        BotName|contains: 'Agentforce'
        InvokingUserId: null
    condition: 1 of them
falsepositives:
    - Legitimate security-research or bug-bounty use of interactsh/OOB-callback domains from an authorized testing host
    - A Slack app-permission change that legitimately anonymizes bot replies for a non-Agentforce integration sharing the same naming convention
level: critical
```

**Known false positives:**

- Authorized security-research or bug-bounty traffic legitimately using interactsh/OOB-callback domains from a sanctioned testing host
- A benign Slack workflow/app sharing the 'Agentforce' naming convention that anonymizes bot replies by design

**Tuning notes:** Maintain and periodically refresh the OOB-interaction-domain allowlist/denylist (oast.fun, interactsh.com and sibling instances rotate); if Slack audit log export doesn't expose InvokingUserId directly, correlate on the absence of a preceding 'confirmation_requested'/'confirmation_granted' audit event pair instead of a null field.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- SalesBleed: Prompt-Injection Payload Submitted via Public Web-to-Lead Form (KQL) (locked: unlock with TL Creds)
- SalesBleed: Agentforce Cross-Object Query Records Triggered by Poisoned Lead Review (KQL) (locked: unlock with TL Creds)
- SalesBleed: Subdomain-Encoded DNS Exfiltration and Anonymous Agentforce Slack Reply Hijack (KQL) (locked: unlock with TL Creds)
- SalesBleed: Prompt-Injection Payload Submitted via Public Web-to-Lead Form (Sigma) (locked: unlock with TL Creds)
- SalesBleed: Agentforce Cross-Object Query Records Triggered by Poisoned Lead Review (Sigma) (locked: unlock with TL Creds)
- SalesBleed: Subdomain-Encoded DNS Exfiltration and Anonymous Agentforce Slack Reply Hijack (Sigma) (locked: unlock with TL Creds)
- SalesBleed: Prompt-Injection Payload Submitted via Public Web-to-Lead Form (SPL) (locked: unlock with TL Creds)
- SalesBleed: Agentforce Cross-Object Query Records Triggered by Poisoned Lead Review (SPL) (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sources

- [SalesBleed: Indirect Prompt Injection and 0-Click Data Exfiltration on Agentforce](https://labs.zenity.io/post/salesbleed-0-click-data-exfiltration-on-agentforce)
- [SalesBleed: Hijacking Agentforce in Slack for Anonymous Phishing](https://labs.zenity.io/post/salesbleed-hijacking-agentforce-in-slack-for-anonymous-phishing)
- [Zero-Click Vulnerabilities in Salesforce Agentforce Expose Wider AI Agent Risk](https://www.infosecurity-magazine.com/news/vulnerabilities-salesforce-ai/)
- [Salesforce Agentforce vulns allowed 0-click CRM data theft, anonymous phishing](https://www.theregister.com/security/2026/09/24/salesforce-agentforce-vulns-allowed-0-click-crm-data-theft-anonymous-phishing/5298958)
- [Salesforce Indirect Prompt Injection Vulnerability Enables 0-click Data Exfiltration](https://cybersecuritynews.com/salesforce-salesbleed-vulnerability/)
- [Salesforce Agentforce Flaw Enables 0-Click Data Exfiltration via Prompt Injection](https://gbhackers.com/salesforce-agentforce-flaw/)
- [Zenity Labs Uncovers SalesBleed, 3 Salesforce Agentforce Flaws Enabling Zero-Click CRM Data Theft and AI Agent Impersonation](https://finance.yahoo.com/technology/ai/articles/zenity-labs-uncovers-salesbleed-3-140200553.html)
- [The lethal trifecta for AI agents: private data, untrusted content, and external communication](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-2710
