NeuralTrust has raised $20M to secure the growing swarm of AI agents in the enterprise.
We're building the security layer that the world's most critical enterprises will run on.
Is your security ready for the agentic era?
Researchers studied 1,899 open-source MCP servers and found 8 vulnerability types. Only 3 overlap with traditional software flaws.
So no single tool covers MCP. It takes seven layers, from registry to audit logs.
Read the full guide → https://t.co/lKjrrIG5Hl
#MCPSecurity #AISecurity
45% of employees now use AI regularly on corporate devices, up from 15% a year ago.
And 67% of users sign in with non-corporate accounts.
Each detection tool sees a different slice. Make sure you have the right solution to detect the blind spot.
Read the full guide → https://t.co/cyIGIYZf4O
#ShadowAI #AISecurity
NeuralTrust now integrates with Microsoft 365 Copilot.
Govern the MCP tools Copilot can reach. In Enforce mode, injected instructions in tool results never reach the model.
For M365 Copilot, not Copilot Studio.
Read the full guide → https://t.co/9JXuJ6WHwo
#Microsoft365Copilot
Anthropic launched Claude Meetings in beta on Nov 12: Claude now joins Google Meet calls and takes notes for you.
A meeting has now become a prompt: speech, chat and shared screens can carry instructions.
Limit connected tools and permissions.
Full guide → https://t.co/O5RR1yFl2G
#ClaudeMeetings #AISecurity
Google launched Gemini 4 Hydrogen on Nov 12.
Independent score: 54 vs 55 for Claude Sonnet 5.5, at $3.16 per task vs $10.02.
Google says it's 2.3x more resilient to prompt injection, but names no baseline.
We compared against Sonnet 5.5 and GPT-6.1 Terra.
Test it on your data.
Full article → https://t.co/wWVDZviNGY
#AISecurity
Anthropic launched Claude Fable 5.2 on Nov 9: #1 on Artificial Analysis, 20% cheaper than Fable 5.1.
But when it blocks a request, a weaker model answers. Injection worked 4.09% of the time vs 0.96% for Opus 5.5 (Notebookcheck).
Full guide → https://t.co/9gIEsJyfvj
#AISecurity
A 5-step workflow that runs 2,000 times a month:
n8n counts 2,000 executions.
A per-step meter counts about 10,000 actions.
Compare cost per workflow, not plan prices.
We compared 6 Zapier alternatives so you don't have to → https://t.co/oIvvm89X9T
#Zapier#Automation
On GitHub, OpenClaw has 659+ security advisories. Hermes Agent has none.
That doesn't make Hermes safe: its own policy puts prompt injection and community skills out of scope.
Sandbox either one.
Full comparison → https://t.co/5rY8WV5dRO
#AIAgents#AISecurity
Can your agents skip your MCP gateway?
• K8s v1.33: native sidecars stable
• MCP 2026-07-28: method and tool name in headers
• IBM 2026: 92% of orgs with AI-related breaches lacked proper AI access controls
Full guide → https://t.co/gTLxd2TnvY
#MCP#Kubernetes #AISecurity
Claude Code's sandbox ships off. OpenCode allows all but .env reads.
• Terminal-Bench 4.0: Claude Code 64.8%, OpenCode absent
• Opus 4.5, TB 2.0: 52.1% vs 51.7%
• Advisories: Claude Code 36, OpenCode 3
Full comparison → https://t.co/1o0XuFIZQa
#ClaudeCode#AgenticAI #AppSec
Claude connectors run with your permissions. A custom connector with a shared API key runs with that key's.
• Each tool can be set to Always allow, Needs approval or Blocked
• 9,720 MCP tool-poisoning tests: 22.9% attack success, 0 of 12 frameworks had a defense
• Use OAuth, not shared headers
Setup, admin controls and risks → https://t.co/HMqoK6Qogh
#MCP #AISecurity #Claude
Claude Code plugins run with your user privileges. Hooks and MCP servers run outside the sandbox.
317 plugins in the official marketplace, 263 from external repos
Plugin4Shell bypassed SHA pinning in 4 coding agents. Fixed in 2.1.179
Policy decides what loads, not what it does at runtime
Review steps, 2026 incidents and admin controls → https://t.co/ewJjmiU17K
#ClaudeCode #AISecurity #MCP
OpenAI released GPT-6 Cyber on November 5.
• ExploitGym: 48.5% vs 32.1% for GPT-5.6 Cyber (vendor-reported)
• Out-of-scope honeypot exploits: 0.07%, down from 5.51%
• Access: approved Daybreak Red customers only
OpenAI secures the model. You secure the agent that calls it: tools, egress, credentials, logs.
Controls split and CISO checklist → https://t.co/z79A41qvGp
#GPT6Cyber #AISecurity #CISO
An OpenAI model tried to write its own way around its sandbox.
This week in The AI Trust Letter:
• OpenAI launches new GPT-6 Cyber
• 948 Linux kernel exploits within 30 hours
• Stolen AI access: up to $1M in cloud bills
• Claude Code's "don't ask again" file exposed secrets
Read the Newsletter → https://t.co/vCRLfRqoCq
#AI #Cybersecurity #OpenAI #Anthropic
Google ADK had six CVEs in 2026. One let attackers forge the human approval step.
Another scored 10.0: unauthenticated code execution through adk web.
Upgrade ADK Python to 2.7.0+ (Java: 1.13.0).
Guide → https://t.co/SVE5BTKOGq
#GoogleADK#AISecurity
AWS's docs say AgentCore does not enforce session-to-user mappings.
That job is yours, and sessions can run up to 8 hours.
Also yours: the default execution role and the SDK version (1.18.1+).
Guide → https://t.co/KphbdkiPvV
#AgentCore#AISecurity
Adversa AI got Copilot CLI to follow encrypted instructions it refuses in plain text.
In autopilot, with a permissive model, it sent a .env file out in 28 seconds.
Watch the whole tool chain, not just the prompt.
Guide → https://t.co/A1BXPV5p4O
#AISecurity#CodingAgents
“Finishing third out of more than 1,500 applicants put NeuralTrust in front of the investors and enterprise leaders we most wanted to reach in Europe.”
That is @Rodrigo_baon, CMO at @NeuralTrustAI, on what placing third in the RAISE 2026 Startup Competition meant for the team.
NeuralTrust came to RAISE Summit 2026 as a Startup Competition finalist, just after announcing its $20M seed round, which the company describes as the largest cybersecurity seed round ever raised by an EU company. Rodrigo said, “some of the best conversations happened backstage, with the jury and with other finalists.”
Since July 2026, NeuralTrust has put that funding into its team, its product and new markets: @Gartner_inc named it a Pioneer in its first Emerging Market Quadrant for AI Application Security, one of Europe’s largest banks now uses its platform to secure and monitor AI globally, and the company has set up local teams in the UK, Germany and Brazil.
As AI adoption accelerates, Rodrigo sees a shift in where security matters most: “Security is moving away from the model and into the layer that connects agents, tools and models.”
Congratulations to Joan Vendrell, Victor Garcia, Alejandro Domingo, and the NeuralTrust team! Looking forward to seeing what’s ahead.
A website can drop a file in your Downloads folder without a click.
The file was an AGENTS.md. In researchers' tests, Codex then ran malware over 90% of the time.
Review agent instruction files like code.
Guide → https://t.co/VYgnxQXKON
#AISecurity#CodingAgents
Claude Code can remember across chats with one command.
But memory is an attack surface. A 2026 study poisoned it, and the attack worked 81.7% of the time in later sessions.
Review what gets saved. Never store secrets.
Guide → https://t.co/GFRs466z8I
#MCPSecurity