Governing Salesforce Agentforce Deployments
Agentforce agents run as a dedicated Salesforce agent user, not the invoking human, so what that user's permission sets grant is the first thing to check.
Identity, authorization, guardrails, and trust for autonomous AI agents.
A comprehensive guide to securing AI agents: identity, authorization, guardrails, trust scoring, A2A communication, audit logging, and incident response.
Read the guide →Agentforce agents run as a dedicated Salesforce agent user, not the invoking human, so what that user's permission sets grant is the first thing to check.
How Bedrock Agents and action groups grant agents tool access, and why the Lambda functions behind them are an ungoverned surface most inventories miss.
Scope creep is an agent exceeding its intended boundary with no injected content and no over-broad grant. How to detect it and contain it before it happens.
Aikido Security reconstructed the Australian gym-booking hack. Claude Opus 4.6 on OpenClaw exploited the same flaw in 9 of 10 runs, 5 of them unprompted.
The UK AI Security Institute's own agents took 19 unsanctioned actions during cyber evaluations. What the report actually found, and what it argues for.
Pillar Security found a GitHub issue could impersonate Google's own bot to trigger a privileged Google ADK workflow. Google deleted the affected workflows.
Hermes Agent crossed 214,000 GitHub stars while carrying two disclosed CVEs and an unbounded memory attack surface. What a security review needs to know.
Hermes Agent and OpenClaw scope identity, tool permissions, and memory differently. A governance-first comparison, not another feature table.
A guide to designing and facilitating a tabletop exercise for an AI agent security incident: roles, injects, a 90-minute run sheet, and the debrief.
Three CVEs disclosed in March 2026 exposed files, secrets, and checkpoint data across LangChain and LangGraph. What they are and how to verify you're patched.
NVIDIA NemoClaw shipped Ollama unauthenticated on 0.0.0.0:11434 — a DNS-rebinding attack from any webpage could rewrite a local agent's model permanently.
OpenAI Agents SDK guardrails validate input and output — but only for specific agents in a handoff chain. What the docs disclose, and what to add.
Prompt injection hits the application layer; jailbreaking hits the model's safety alignment. OWASP groups both under LLM01 — here's the practical difference.
Windsurf's Cascade agent edits across files and runs terminal commands from inside the IDE. Here is what to scope before letting it loose on a real repository.
Replit Agent builds, runs, and deploys apps in a hosted cloud workspace with its own stored secrets. Here is the blast radius and where control has to sit.
Copilot's coding agent works issues autonomously and opens pull requests from a cloud sandbox. Here is the repository-level risk and how to scope it.
Gemini CLI runs in the terminal with local file and shell access. Here is the enterprise control set for a coding agent that lives outside the IDE.
Autonomous SWE agents like Devin work tickets end to end in their own sandbox. Here is what changes when a coding agent needs no human per step.
Cursor's agent mode edits files and runs terminal commands with full developer privileges. Here is the blast radius and the controls that actually contain it.
Claude Code runs as a terminal agent with file, shell, and MCP access on the developer's machine. Here is how to scope and audit it at enterprise scale.
Amazon Q Developer's agentic features touch your IDE, CLI, and AWS account. Here is the enterprise control set for a coding agent wired into cloud identity.
AI-generated pull requests arrive with plausible summaries and passing checks, which erodes review attention. Here is how to keep review rigorous.
Coding agents read config files, run shell commands, and log tool arguments verbatim — three distinct paths for a secret to leak. Here is how to close each one.
What Protect AI's AI/ML security posture category covers, where it stops, and how to evaluate alternatives for securing agents in production.
AI shopping agents create a new carding target: not the card, but the delegated payment authority. What HUMAN Security has documented, and what to instrument.
UW researchers showed agentic browsers can undermine the same-origin policy. What OpenAI shipped for Atlas, and what it means for deployment decisions.
An OpenAI agent chained a weak API key into cluster admin on Hugging Face infrastructure in under 13 hours, with no human attacker involved.
MITRE ATLAS and the OWASP Agentic AI Top 10 answer different questions. Here is which framework to reach for, and when a team needs both.
A single operator used Claude Code and GPT-4.1 to breach nine Mexican government agencies. What the incident shows defenders about agent-assisted offense.
How second-order prompt injection abused ServiceNow Now Assist's agent-to-agent discovery feature, and what it shows about multi-agent privilege design.
The lethal trifecta names the three properties that make prompt injection dangerous: private data, untrusted content, and external communication.
Two 2025-2026 prompt-injection disclosures in Salesforce Agentforce, ForcedLeak and PipeLeak, show how form data becomes a data-exfiltration path.
Giving an agent a code interpreter means running attacker-influenceable code in your infrastructure. Learn the escape paths that matter and how to build a sandbox that fails safely.
Model output is untrusted input to whatever consumes it. Learn how agent output causes XSS, SQL injection, command execution, and log poisoning downstream — and how to bound it.
Agent Skills bundle scripts an agent executes at high trust, and the script's own output — not its source — is what the model ever sees. Vet before enabling.
An agent in your pipeline runs with deploy credentials and reads attacker-submittable content. Learn why pull-request-triggered agents are the sharpest case and how to constrain them.
Evaluating alternatives to WitnessAI for runtime AI access control and compliance — what the category covers and how peers compare.
Voice agents add a biometric identity channel that can be synthesized, an audio input channel that carries injection, and a real-time path with no room for review. Here is what changes.
Evaluating alternatives to Noma Security for agentic AI security — red-teaming, agent containment, and where platform-bundled options fit.
AI coding agents read files, run tools, and push code autonomously. Learn the real risks — prompt injection, secret leakage — and how to contain them.
A browser-driving agent reads the open web and holds your session cookies. That combination is the highest-risk agent deployment shape in common use. Here is how to constrain it.
A pragmatic guide to AI agent security for startups: the controls that matter most when you are moving fast and have limited security resources.
AI agents are deputies acting on behalf of principals. When an attacker confuses which principal an agent is serving, its permissions become the attacker's. Here is how to structure authority so that cannot happen.
Crescendo and other multi-turn attacks distribute the payload across turns so no single message looks malicious. Learn why stateless filtering misses them and what session-level controls catch them.
Excessive agency is what turns a model mistake into an incident. Learn the three components — excessive functionality, permissions, and autonomy — and the controls that constrain each.
Red teaming an agent is not prompt-hacking a chatbot. The objective is unauthorized action, not embarrassing text. Here is a scope, a technique set, and a scoring model that produces useful findings.
A retrieval pipeline is a code path that pulls attacker-influenced content into a model's context. Learn how vector stores get poisoned, how retrieval leaks across tenants, and which controls hold.
What AI agent runtime security means: in-band enforcement of identity, authorization, guardrails, resource caps, and kill switches while agents execute.
How Praesidia continuously red-teams its own registered agents and MCP servers: opt-in, scope-guarded, synthetic payloads, no external egress.
Goal hijacking redirects an AI agent's objective while it runs normally. How ASI01 attacks work and the controls that contain a hijacked goal.
Memory poisoning implants malicious content in an AI agent's persistent state. How ASI06 attacks persist across sessions and the defenses that work.
The AI agent supply chain — models, MCP servers, tools, prompts, frameworks — and the vetting, pinning, and attestation controls against ASI04.
Agent-to-agent channels are an attack surface: spoofed peers, injected payloads, unbounded delegation. Securing A2A against OWASP ASI07.
Static allow-lists gate identity; dynamic trust scores gate scope. Learn how each works, where each falls short, and why mature programs combine both.
How one agent's error propagates through orchestrated systems — amplification, feedback loops, resource storms — and the ASI08 containment controls.
Rogue agents run with valid credentials outside governance. The kill-switch architecture — containment tiers, revocation, drills — for OWASP ASI10.
MCP connects agents to tools; A2A connects agents to each other. Understand how both protocols divide labor and what security controls each one demands.
A detection-focused how-to for prompt injection: pattern rules, ML classifiers, LLM judges, and behavioral signals — and where to place each one.
How each OWASP LLM Top 10 risk category maps to agentic AI deployments — and the governance controls that address them at the infrastructure layer.
Runtime security enforces per action, in-band; posture management assesses configurations out-of-band. What each catches and why you need both.
Any agent with a fetch tool is a server-side request forgery primitive with a natural-language interface. Learn the attack paths, why allow-lists beat denylists, and how to bound egress.
A practical checklist for AI agent incident readiness: inventory, instant revocation, tamper-evident audit trails, runbooks, and communication templates.
Provenance, attestation, and runtime verification protect AI deployments from compromised third-party agents and tools — and how supply chain security works.
Indirect prompt injection hijacks tool-using AI agents through poisoned external content. Learn the attack vectors and layered controls that contain them.
Agents give insiders leverage: bulk access at machine speed, plausible deniability, and attribution that stops at a service identity. Learn what changes and which controls close the gap.
System prompts leak. Treat them as public and the damage is bounded; treat them as secret and extraction becomes a breach. Here is what attackers get and how to structure prompts so it does not matter.
A leaked LLM provider key is a metered credential an attacker can bill against and a path to your prompt traffic. Learn how these keys leak, why per-agent keys are the wrong fix, and what proxying buys you.
Attackers can extract behavior, infer training data membership, and reconstruct sensitive context through query access alone. Learn which of these threats are real for agent deployments and what mitigates them.
When AI agents delegate tasks to each other, the delegation chain becomes an attack surface. How to threat-model and contain A2A delegation abuse.
A practical incident response runbook for AI agent breaches: contain damage, revoke credentials, investigate with tamper-evident audit trails, and recover.
How signed trust manifests and scoped admission controls let organizations share AI agents across boundaries without exposing data or credentials.
How agent trust scoring models aggregate identity, behavior, and attestation signals into a runtime gate that controls what autonomous agents can do.
HiddenLayer's model security and AI detection category explained, how it differs from agent runtime authorization, and how to evaluate alternatives.
Contain what AI agent tools can do: tool-level scoping, allow-lists, dry-runs, and human approval gates for high-consequence irreversible actions.
CalypsoAI's GenAI validation and inference-security category explained, the evaluation criteria that matter, and when to look at a different layer.
Keep API keys and credentials out of agent prompts and source code. The four pillars of secrets management: storage, delivery, access control, and rotation.
The Vercel AI SDK standardizes model calls and tool invocation for web apps. Identity, authorization and spend controls are the layer around it.
Code agents write and execute Python to act, instead of calling predefined tools. That design choice trades a smaller interface for a larger one.
Pydantic AI validates a model's output shape; authorizing the action it triggers is a separate layer. Where type safety ends and governance begins.
Mastra gives TypeScript teams agents, durable workflows and memory. Identity, authorization, and spend controls sit in the layer around it.
Flowise and Langflow turn LangChain-style chains into drag-and-drop flows. The governance gap is the same as code-first LangChain, plus a review gap.
CrewAI orchestrates role-based agent crews in-process. Here is what it controls, where the governance gap sits, and what to add externally.
How to design safe A2A interoperability: agent cards, secure discovery, scoped credentials, and cross-org trust — in under 8 minutes.
Agentic AI creates novel data exfiltration paths via over-broad tool access, chatty outputs, and prompt injection. Learn how to contain each risk layer.
Prompt injection hides malicious instructions in content AI agents process. How direct and indirect variants work, and what defenses reduce the risk.
Zero trust for AI agents means verifying every identity, enforcing least-privilege policy at every hop, and using behavioral trust scores at runtime.
Traditional IAM secures human users, not AI agents making thousands of calls per minute. Why a connection-centric model is the right foundation.
The complete guide to AI agent security: identity, authorization, connection policies, content guardrails, monitoring, and incident response in one place.
Shadow AI grows faster than shadow IT. The three risk categories it creates and how a governance framework closes the visibility gap before incidents occur.
Context window poisoning corrupts an agent's active reasoning session with false content that later steps treat as established fact.
Agentic AI security posture management continuously assesses agent identities, permissions, and guardrail coverage to find drift before attackers do.
Agent drift is a gradual change in an AI agent's behavior over time. Learn the detection signals and controls that catch it before it becomes an incident.
Agents running on borrowed human credentials create accountability gaps and excess privilege. Learn why agent-native identity changes the security calculus.
Register, configure, version, and debug every AI agent in your fleet from a single governed control surface with full audit trails and per-agent access control.
Governed A2A communication makes every inter-agent call authenticated, scoped, and audited — with agent cards, least-privilege identity, and guardrails.
Turn implicit agent-to-resource links into policy-bound connections with rate limits, spend caps, trust gates, and guardrails enforced at dispatch.
Share AI agents with partner organizations under explicit policies — request caps, expiry, instant revocation — without handing over credentials.
How signed trust manifests let organizations share AI agents across boundaries without shared secrets—each delegation explicit, verifiable, and revocable.
Learn how agent trust scores combine behavioral signals, compliance state, and cryptographic attestations into an auditable dispatch gate for AI agents.