Upgrading from 4.x? 5.x needs Node 22.14+ or 24+; Node 20 and 23 are unsupported. On Node 20, npm warns and still installs unless engine-strict is set; the database engine will not load. Read the guide.
Blog
Security insights, product updates, and AI agent protection guides.
The Prompt Was Clean. The Model Was Not.
Microsoft research shows a coding agent can carry a backdoor, trained in after pre-training, that turns an ordinary request into an attacker-chosen tool call; input filters may see nothing, so controls belong at the tool-execution boundary.
Read articleRunning a Reverse-Engineering Agent Safely: REA and ShieldCortex
REA connects agents to native-binary analysis through Hopper, Ghidra or IDA, plus separate tools for apps, websites and saved captures. Isolate the session, pin a version and review the setup plan; ShieldCortex adds memory-write scanning, skill scanning and an optional tool gate on supported hosts.
ShieldCortex 5.13.0: Doctor Warns About Stale ShieldCortex Copies
A new read-only shieldcortex doctor row warns about ShieldCortex copies at 5.2.1 or older in fixed install locations, which can share the chained audit database and still add unchained rows and checkpoint-less deletes that shieldcortex ledger verify reports. Minor on 5.12.0 with no new configuration; Action Guard stays off by default.
ShieldCortex 5.12.0: Embedding Model Cache Settings, Config Re-sign, Dashboard Update Check Switch
Opt-in settings let the embedding model load from a relocated cache, run offline from a pre-seeded cache or download from a validated mirror; this covers the embedding model only and is not a product-wide offline mode. A reviewed config.json that fails its signature check can be re-signed, and the dashboard update check can be turned off. Minor on 5.11.0; Action Guard stays off by default.
ShieldCortex 5.11.0: Config Tamper Evidence, Clearer Watch Only, Session-End Timeout Digest
A write on a config.json that fails its signature check now refuses instead of re-signing it, Watch only stops denying plain read-only inspection of ShieldCortex's own config on headless sessions and names its signals, and the OpenClaw plugin can send one report-only digest of recorded approval timeouts at session end. Minor on 5.10.7; Action Guard stays off by default.
ShieldCortex 5.10.0: One Command for Storage and Memory Hygiene
A new shieldcortex maintain command consolidates short-term memory over its cap, backfills missing vectors, vacuums when at least 20% of database pages are free, then runs doctor's Memories, Semantic coverage and Disk checks, exiting 1 while a FAIL remains. Plain maintain can delete and evict memories, so run --dry-run first. Minor on 5.9.0 with no default changes.
ShieldCortex 5.9.0: Fewer False Alarms, Better Global Install Coverage
Action Guard again allows narrowly defined data-reading heredocs and adds recognised written argv-list global installs to the approval gate when on and enforcing. The session lease skips read-only global npm queries while keeping redirection-attached installs in scope. Action Guard stays off by default in the shared ShieldCortex config.
ShieldCortex 5.8.0: Recall That Answers What You Asked, Action Guard for Claude SDK Browser and Computer Use, Recall Repetition Check
Prompt recall now queries the owner's message instead of the OpenClaw envelope. The new Claude SDK browser and computer toolsets export observes calls by default, and doctor warns when recent interactive recalls repeat the same candidate set. Vector recall remains off by default.
ShieldCortex 5.7.0: Bulk Cloud Sync Redaction, Recall Lanes, Clearer OpenClaw Setup
A minor release on 5.6.0. The dashboard's bulk quarantine sync now applies the same filters and PII and credential redaction as the automatic sync, prompt-time recall in the Claude Code UserPromptSubmit hook is scoped by lane with a new PERSONAL tier, and setup openclaw reports every plugin install refusal. Action Guard stays off by default.
ShieldCortex 5.6.0: Plain-English Approval Cards, Maintenance Before Dream Mode, Safe --help
A minor release on 5.5.0. Approval cards say what, why and who, consolidate runs memory maintenance before Dream Mode and can remove short- and long-term memories, and install --help, setup --help and doctor --help now print usage without installing or running checks. Action Guard stays off by default.
The Vault Kept the Secret. The Agent’s Shell Read It Anyway.
Unit 42 showed an injected support ticket steering a default-configured AWS AgentCore Harness agent into lifting a vault-held service credential from runtime memory. Vaults protect at rest and in transit, not in use. Scope tools per invocation, shrink service accounts, watch egress.
Your AI Agent Is a Cloud Identity. Its Permissions Are Only Half the Story
Unit 42 maps cloud identities from behaviour, not names. For AI agents a third question matters: what context caused the action? IAM is not enough. ShieldCortex sits at the context boundary — not as CloudTrail.
When Your Agent's API Key Becomes Someone Else's Product
Unit 42 documented stolen AI platform API keys being fed into gray-market transfer stations within minutes, running up nearly a million dollars in usage. Agents that remember keys turn a rotation ticket into inventory.
When the Attacker Runs an Agent
Unit 42 documented a Chinese-speaking threat actor running an autonomous agent that enumerated targets, sourced exploits, and launched attacks with limited human input. Agentic AI is now an offensive operator — and the discipline that contains it is the same discipline that secures your own agents.
When Your AI Agent Trusts a Domain That Does Not Exist
Unit 42’s phantom-squatting research shows LLMs invent plausible brand domains that attackers register. Once agents can fetch, install, call tools, or write URLs into memory, hallucination becomes a supply-chain decision.
Your AI Agent’s Skill Marketplace Is Now a Supply Chain
Third-party AI-agent skills are becoming the new plugin economy: easy to install, powerful by design, and dangerously close to credentials, files, shell commands, and long-term memory. Scanner verdicts help, but they are not a trust boundary.
When AI Agents Enter CI/CD, Prompt Injection Becomes a Secrets Problem
Microsoft Threat Intelligence disclosed a prompt injection pathway in the Claude Code GitHub Action that could reach workflow secrets under specific conditions. Once an agent reads untrusted content and runs near credentials, prompt injection stops being a chat problem and becomes an infrastructure security boundary.
When Old Dependencies Become New Backdoors: Lessons from the Laravel Lang Supply Chain Attack
Socket Security reported 700+ historical versions of Laravel Lang packages poisoned with an RCE backdoor and a credential harvester. Pinned lockfiles, cached CI artifacts, and old Docker images are all in scope — and AI coding agents make the blast radius worse.
Proactive Recall: Why AI Agents Should Remember Before They Reply
Historical post (April 2026). Automatic inject is now off. Native host memory stays the brain; ShieldCortex scans writes and gates tools.
Introducing Cortex — Your Agent Learns From Its Mistakes
AI agents repeat the same mistakes because they don't learn between sessions. Cortex captures what went wrong, runs pre-flight checks before tasks, and graduates rules once mastered. It stores those rules locally.
NVIDIA Chose OpenClaw. Here's How We Secure It.
NVIDIA released NemoClaw — OS-level sandboxing built on OpenClaw. They called it 'the operating system for personal AI.' Here's what NemoClaw does, what it doesn't, and where ShieldCortex fills the gap.
Why Runtime Security Isn't Enough — The Case for Memory Integrity
IronClaw secures the runtime. But what secures the memory? Even the most locked-down agent sandbox can't protect against poisoned persistent memory. Here's why you need both runtime security and memory integrity.
Introducing Iron Dome — Behaviour Protection for AI Agents
Your defence pipeline stops poisoned inputs. Iron Dome stops poisoned outputs. Control what your AI agent can do — not just what it remembers.
ShieldCortex Now Speaks Python — Protect AI Memory from CrewAI, LangChain, and Beyond
The official Python SDK is live on PyPI. Scan AI agent memory for prompt injection and credential leaks — with built-in CrewAI and LangChain integrations.
Real-time LLM Scanning for OpenClaw — Defence at the Pipeline Level
The new ShieldCortex plugin for OpenClaw scans every LLM input for threats and auto-extracts memories from outputs. Fire-and-forget — never blocks your agent.
MCP Security: What Claude Code Users Need to Know
The Model Context Protocol is powerful, but MCP servers have direct access to your system. Here's how to protect yourself.
The 5 Ways Hackers Can Poison Your AI Agent's Memory
AI agents with persistent memory are vulnerable to a new class of attacks. Here are the 5 techniques attackers use — direct injection, encoded payloads, fragmentation, context manipulation, and trust escalation.
Introducing ShieldCortex: The Security Layer Your AI Agent is Missing
AI agents are everywhere. They're writing code, managing infrastructure, processing emails. And increasingly, they have persistent memory. This is powerful. It's also dangerous.