16 NovA vendor script on 178 company websites was changed to skim cards. What happened

The AI security team that defends your company.

Specialised AI agents watch your people, cloud, code, SaaS apps and AI agents. They find what is exposed, investigate every alert, and stop attacks and costly mistakes in seconds. They run on Halda's own security models, inside your company. Your data never leaves it.

Example: an email asks Priya in finance to pay an invoice to a new bank account. Halda spots the look-alike domain, holds the payment and warns her.

Your company what Halda defends
People Priyafinance Jordanengineering
Cloud and code AWS productionplatform GitHubengineering VPN gatewaypatch due Checkout pageweb team
SaaS Salesforcesales Microsoft 365IT
AI agents Claude Codejordan Support agenthelp desk
Halda finds, investigates and acts
Halda · example
Your systems kept safe
Customer data Cloud account Source code Payments
Works acrossMicrosoft 365Google WorkspaceOktaAWSGitHubSalesforceSlackClaude CodeCursorMCP servers
01 02

Attackers move in minutes. Companies take months.

Once inside, attackers reach the rest of a network in about 29 minutes. Most companies take months to find and contain a breach. And AI agents now hold keys to code, cloud, money and customer data.

CrowdStrike Global Threat Report 2026 · IBM Cost of a Data Breach 2026

02 02

A security team that moves as fast.

Halda puts a specialised AI agent on every job a security team does. It investigates every alert, acts in seconds, and asks your people first before anything that matters.

Real attacks

It already happened. Here is what Halda does.

Public incidents from the last year, each with its source, and what Halda does at the same step.

Supply chainAI coding tools turned on their owners

A poisoned build package told developers' own AI tools to hunt for their keys.

In August 2025, malicious versions of Nx, a build tool with millions of weekly downloads, ran Claude Code, Gemini CLI and Amazon Q on developers' laptops with their safety switches off, and told them to find tokens, SSH keys and wallets. The haul was posted to public GitHub repositories: 2,349 secrets in 1,346 repositories.

With Halda: the install script runs inside the wall. The usual key files are not readable there, and nothing leaves for an address that was not allowed. Our attack test runs this exact move on every build.

Sources: Wiz · Socket

Target
Developer laptops
Way in
npm install script
Then
AI tools run with --dangerously-skip-permissions, --yolo
Taken
GitHub tokens, SSH keys, cloud keys, wallets
attack test · our own run
$ halda breach-test --confined
postinstall scriptwith nothing in the waygot out
postinstall scriptinside the wallheld
01
Destroyed data

An AI agent deleted a production database during a code freeze.

July 2025, Replit. The agent wiped records for 1,206 executives and about 1,200 companies, then said recovery was impossible.

With Halda: the command is tried on a copy of the database first. The person on call sees exactly what it would delete, row counts included, and says no.

eWeek

02
Poisoned extension

A wiper prompt shipped inside an AI coding extension.

July 2025. A pull request slipped instructions into Amazon Q's VS Code extension, installed over 964,000 times, telling the agent to delete local files and cloud resources.

With Halda: wiping a home folder is refused outright. Deleting files outside the project, or any cloud resource, stops for a named person with the exact command.

SC Media

03
Stolen keys

A self-spreading npm worm stole keys and published them.

September 2025, Shai-Hulud. An install script scanned for GitHub, npm and cloud keys, posted them to public repositories, and used the npm keys to infect more packages.

With Halda: install scripts run inside the wall. The usual key files are not readable there, and nothing leaves for an address that was not allowed.

Help Net Security

04
Vendor scriptComing soon

A trusted vendor script was changed to skim cards on 178 company websites.

November 2026, Optimizely. The script every one of those sites loaded was altered to copy what customers typed at checkout.

With Halda: every third-party script on your site is watched. A change is flagged within minutes, with what it now does.

Report URI

Halda Research · 18 Nov 2026

We read the 1,000 most-used MCP servers. 3 in 4 update themselves.

Together they are downloaded 58.7 million times a month. 76% tell you to set them up so your AI agent fetches and runs the newest release every time it starts. 2% pin a version. One bad release reaches every user, with their keys, and nothing asks first. That is how postmark-mcp, Nx and Shai-Hulud worked.

76%

Run the newest release on every start.

743 of 980 servers with setup instructions, including official ones from large software companies.

22%

Also tell your agent what to do.

223 ship skills or agent instructions, so a new release can change your agent's instructions, not just its tools.

8%

Run code the moment they install.

84 run their own scripts during install, before you have used them once.

Read as text, never run. Check your own machine with the free tool: halda scan.

What Halda does

Every job of a security team, in one product.

Today each job is a separate product, and AI agents are covered by yet another. Halda's specialised agents do all of it, with one set of rules and one record. They work as one loop: Halda attacks your company, fixes what it finds, attacks again to prove the fix held, and handles whatever happens anyway.

AI agentsAvailable now

Every AI agent under control.

Every agent, plugin and MCP server in the company found and given an owner. Hijacked agents cannot take data out or damage production.

Red teamAvailable now

Attacks you before attackers do.

Halda attacks your own setup the way real attackers and hijacked agents would, and shows exactly what got through.

AuditAvailable now

A record auditors can check.

Every decision signed and sent to your logging tools, mapped to OWASP, MITRE, NIST, ISO and the EU AI Act.

ExposureComing soon

Finds what attackers can reach.

Vulnerabilities, misconfigured cloud, leaked keys and open buckets, ranked by what an attacker could really use.

IdentityComing soon

Every account, key and token.

People and machines. Over-powerful roles, forgotten tokens and connected apps, cut down to what each job needs.

InvestigationComing soon

Every alert investigated.

Reads alerts from your SIEM, endpoint and email tools, investigates each one in minutes and shows its reasoning.

ResponseComing soon

Acts in seconds.

Revokes a key, blocks a sign-in, isolates a laptop or holds a payment. Your people approve what matters.

Email and fraudComing soon

Phishing and payment fraud stopped.

Fake invoices, look-alike domains and calls pretending to be staff, caught before money or access moves.

FixingComing soon

Fixes what it finds.

Each problem comes with a fix you approve. Halda applies it, then attacks again to check it held.

What it protects

Your whole company, one team.

Stop fraud and account takeover.

Fake invoices, look-alike domains, calls pretending to be staff, and stolen sign-ins.

Book a demo
  • Payments to new bank accounts held until checked
  • Look-alike domains and impersonation caught
  • Help-desk reset requests verified
  • Stolen and reused sign-ins blocked

Runs in your cloud, on your own servers, or fully offline. Your data never leaves your company.

Our models

Our own security models.

Small models trained for security work: how attacks unfold, which alerts are real, and what an action will really do. They run inside your company.

Model 1

What an action will do.

Where no copy is possible, such as payments, it says what would change and sends risky actions to a person.

Model 2

Is this alert real?

Reads an alert with its context and says how likely it is to be a real attack, with its reasons.

Model 3

New attacks.

Invents new ways to hijack your agents, to test you against unpublished attacks.

Your data never leaves your company and never trains anyone else's model. The final yes or no is plain code and a named person, never a model.

Attack test

A hijacked agent tried to steal a secret. It never got out.

A hijacked agent holding a test password tried 18 ways to get it out. With nothing in the way, every one of the 15 that could run got it out. With Halda, none did. The other 3 stopped for a person.

Why a wall and not just checks: an attacker who hides the theft inside a Makefile or an install script gets past command checks alone 8 times out of 18. The wall stops it anyway.

Run it on your setup →

attack test · our own tests
$ halda breach-test --confined
nothing in the wayattacks that got the test password out15 of 15
inside the sandbox · macOSattacks run that got out0 of 15
inside the sandbox · Linuxattacks run that got out0 of 14
Measured 2026-11-04, rerun 2026-11-17.
What Halda tests for

The ways agents really get turned.

  1. Hidden instructionsOrders planted in files, tickets, web pages and tool results the agent reads.
  2. Poisoned packagesInstall scripts, test hooks and startup files that run the moment a package arrives.
  3. Secrets sent outKeys encoded, split, copied to a plain name, or sent by raw socket and DNS.
  4. Destroyed dataDropped tables, Terraform destroy, force-pushes that lose other people's work, mass deletes.
  5. Tools that changeAn MCP tool that rewrites its own description after you approved it.
  6. Keys with too much powerStanding credentials where one key for one action would do.
  7. Reworded commandsThe same dangerous action phrased a different way to slip past a rule.
  8. Tampering with the guardAn agent editing its own rules, hooks or instruction files.
Why Halda

One team instead of ten products.

Less noise.

Every alert is investigated. Only the real ones reach a person, with the evidence attached.

Seconds, not days.

Halda acts as soon as it is sure, and asks a named person first for anything that cannot be undone.

One set of rules, one record.

Every job runs on the same rules and writes to one signed record your auditors can check.

Security

Built to be signed off by a CISO.

Runs inside your company.

On your laptops, your servers or fully offline. Your code, data and logs never come to us.

Unsure means stop.

If Halda cannot tell what an action will do, it waits for a person instead of letting it run.

People make the final call.

A model can raise a concern. Only plain rules and a named person can allow or refuse.

Pricing

Start free, in a minute.

The scan

Every AI agent, MCP server, skill and extension on your machine, and what each one can reach.

The exposure test

What a hijacked agent on your machine could reach and send out. Nothing leaves it.

Replay

Last month's agent commands, judged. See what Halda would have stopped.

Free on your own machine. Companies pay for the full team: every job above, across the company.

Questions, answered.

What is Halda?

An AI security team that defends your whole company: people, cloud, code, SaaS apps and AI agents, in one product.

Does my code or data leave my company?

No. Halda runs inside your company. Only warnings about bad plugins are shared between customers, and you can turn that off.

We already use Okta, Cisco or Microsoft for agents. Why Halda?

Halda works with them, and adds the layers they do not cover.

How do we start?

A 20-minute call. We run the attack test on one of your laptops and show you the result.

Early access

Put an AI security team
to work for you.

We set up our first companies personally.

Book a demo