AlpacaX

The AlpacaX blog—security, product thinking, and what we're learning as we build AI agent execution control.

Product update: what your AI agents did, and why their commands were stopped

Product

Product update: what your AI agents did, and why their commands were stopped

See what your AI agents did this month and how often their commands were refused, all on one dashboard. People and agents both learn why a command was stopped.

Taeyeong Baek
Taeyeong BaekGTM Associate · 10 November 2026
What are your vendors' AI agents running on your servers?

Insights

What are your vendors' AI agents running on your servers?

Their agents can connect with the keys you gave the vendor. Check what ran from your own record, not the vendor's word.

Taeyeong Baek
Taeyeong BaekGTM Associate · 9 November 2026
AI agent policy enforcement: where the policy actually runs

Insights

AI agent policy enforcement: where the policy actually runs

What the enforcement point sees, who can reach it, and what it does when the decision doesn't arrive.

Marco Kwak
Marco KwakHead of GTM · 30 October 2026
Your first Work Session: purpose, approval, and the record

Product

Your first Work Session: purpose, approval, and the record

Four fields and approval where your workspace requires it: how a new Alpacon workspace starts governed work.

Taeyeong Baek
Taeyeong BaekGTM Associate · 29 October 2026
What is AI agent security? Five layers, five questions

Insights

What is AI agent security? Five layers, five questions

Model, identity, policy, execution, monitoring: one question each, and none answers another's.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 28 October 2026
Does Claude Code or Cursor make you an EU AI Act provider?

Insights

Does Claude Code or Cursor make you an EU AI Act provider?

Writing code with an AI agent is not an Annex III use. Ranking engineers with one likely is.

Jungyeon Lee
Jungyeon LeeContent Marketer · 27 October 2026
Product update: intuitive access rules, logins worth a look

Product

Product update: intuitive access rules, logins worth a look

Set access rules by completing a sentence and test them first, see which rule decided a request, and review only the direct logins that need a look.

Taeyeong Baek
Taeyeong BaekGTM Associate · 27 October 2026
Can you prove what your AI agents did on production servers?

Insights

Can you prove what your AI agents did on production servers?

What it could reach, what it was trying to do, what it actually ran, and who checked it. Each needs a record.

Taeyeong Baek
Taeyeong BaekGTM Associate · 26 October 2026
How we patched php2shell in six minutes with an AI agent

Incident

How we patched php2shell in six minutes with an AI agent

What our agent ran, why each privileged command was allowed, and the fixed builds, restart checks and origin lockdown to do it yourself.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 21 October 2026
What Spain's AEPD actually published about an AI agent breach

Insights

What Spain's AEPD actually published about an AI agent breach

AEPD's own account: four phases, chained together, with key details left out.

Marco Kwak
Marco KwakHead of GTM · 16 October 2026
The model you approved isn't the model running today

Insights

The model you approved isn't the model running today

Unless you're pinned to a fixed snapshot, a vendor retrain can land on their schedule, after your sign-off.

Jungyeon Lee
Jungyeon LeeContent Marketer · 15 October 2026
Loopjacking: what you approved isn't what ran

Insights

Loopjacking: what you approved isn't what ran

A researcher reproduced the bypass in tested setups across three released agent frameworks.

Marco Kwak
Marco KwakHead of GTM · 14 October 2026
Product update: how a request arrived, and what reaches you

Product

Product update: how a request arrived, and what reaches you

Approval cards show how a request arrived and whether it was verified, alerts can be muted per server, and extending a session takes a reason.

Taeyeong Baek
Taeyeong BaekGTM Associate · 13 October 2026
What is agentic AI? The definition that actually holds up

Insights

What is agentic AI? The definition that actually holds up

The three properties that define agentic AI are why the label keeps getting abused.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 12 October 2026
Inside the AI agent mass exploitation of PaperCut NG/MF

Incident

Inside the AI agent mass exploitation of PaperCut NG/MF

395 organizations breached by parallel AI agents—the eleventh, not the first, fell in 26 seconds.

Jungyeon Lee
Jungyeon LeeContent Marketer · 9 October 2026
Two CrowdStrike failures, one kernel-mode security risk

Incident

Two CrowdStrike failures, one kernel-mode security risk

CrowdStrike patched a privilege-escalation bug in 12 days. The lesson is where the privilege lived.

Marco Kwak
Marco KwakHead of GTM · 8 October 2026
MCP security can't stop at the install review

Incident

MCP security can't stop at the install review

It passed two clean tool calls, then turned malicious on the third—no one-time review catches that.

Jungyeon Lee
Jungyeon LeeContent Marketer · 7 October 2026
What AI agent access control actually has to check

Insights

What AI agent access control actually has to check

A support bot's admin tools got exploited—not by tricking the model, but by exploiting the access model.

Marco Kwak
Marco KwakHead of GTM · 6 October 2026
Stay on Supabase and Vercel, or run your own servers?

Insights

Stay on Supabase and Vercel, or run your own servers?

Cost makes you look. Ownership makes you consider moving. What usually stops you is the same thing: no one owns infrastructure. That is starting to change.

Taeyeong Baek
Taeyeong BaekGTM Associate · 6 October 2026
Inside JadePuffer, the first documented AI agent ransomware attack

Incident

Inside JadePuffer, the first documented AI agent ransomware attack

A human picked the target. Everything after that was the agent's own call—twice, three weeks apart.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 5 October 2026
How BNOW fixes production without a DevOps team

Testimonial

How BNOW fixes production without a DevOps team

BNOW has no one dedicated to DevOps. Now it hands production work to an AI agent and fixes problems itself.

Taeyeong Baek
Taeyeong BaekGTM Associate · 28 September 2026
A human caught the AI agent. That's not a control you can rely on

Incident

A human caught the AI agent. That's not a control you can rely on

UK AISI's report: human vigilance held. A technical barrier that would reliably hold was never there.

Marco Kwak
Marco KwakHead of GTM · 24 September 2026
AI agent framework harvested thousands of credentials in under 6 hours

Incident

AI agent framework harvested thousands of credentials in under 6 hours

The framework did nothing a skilled attacker couldn't. It just ran without waiting for a human decision.

Jungyeon Lee
Jungyeon LeeContent Marketer · 23 September 2026
Bank regulators wrote AI agents out of model-risk review

Insights

Bank regulators wrote AI agents out of model-risk review

SR 26-2 excludes agentic AI from formal bank model-risk scope, leaving runtime governance to each bank's own risk program.

Marco Kwak
Marco KwakHead of GTM · 22 September 2026
MCP least privilege has two layers, not one

Insights

MCP least privilege has two layers, not one

Which tools a client registers and what judges the commands they run are two different controls.

Jungyeon Lee
Jungyeon LeeContent Marketer · 22 September 2026
Product update: what gets re-checked, and what gets written down

Product

Product update: what gets re-checked, and what gets written down

A token's reach follows its owner's current permissions, revoking a sign-in closes the terminals it opened, and the workspace writes up each period's activity.

Taeyeong Baek
Taeyeong BaekGTM Associate · 22 September 2026
A token that can reach your API isn't cleared to run any command

Insights

A token that can reach your API isn't cleared to run any command

A token has no session and no declared purpose. It has a scope, and a scope can't tell you a command is safe.

Marco Kwak
Marco KwakHead of GTM · 21 September 2026
Zero standing privilege: a 7-question audit for your stack

Insights

Zero standing privilege: a 7-question audit for your stack

Seven yes/no checks to run against your own stack before you trust a vendor's zero-standing-privilege claim.

Jungyeon Lee
Jungyeon LeeContent Marketer · 18 September 2026
MFA fatigue is a scoping problem, not a user problem

Insights

MFA fatigue is a scoping problem, not a user problem

Prompt-bombing wins when every extra MFA check trains the team to auto-approve the one that matters.

Marco Kwak
Marco KwakHead of GTM · 17 September 2026
AI compliance software doesn't solve contractor AI agent risk

Insights

AI compliance software doesn't solve contractor AI agent risk

Every framework requires an identifiable actor. Contractor BYOD and AI agents are what breaks it.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 16 September 2026
A CVSS 10.0 bug turned Metabase's password reset into full admin

Incident

A CVSS 10.0 bug turned Metabase's password reset into full admin

CVE-2026-72898 let anyone reach full admin with zero credentials, then everything Metabase touched.

Jungyeon Lee
Jungyeon LeeContent Marketer · 15 September 2026
Identity mapping in PAM: four handoffs, one fails silently

Insights

Identity mapping in PAM: four handoffs, one fails silently

A login claim refreshes at login. A leaver never logs in again—so nothing ever re-checks, and nothing gets revoked.

Marco Kwak
Marco KwakHead of GTM · 14 September 2026
AI vendor questionnaire checklist: 6 runtime questions to score

Insights

AI vendor questionnaire checklist: 6 runtime questions to score

The six runtime questions to run this week, scored so a failure can't average into an approve.

Marco Kwak
Marco KwakHead of GTM · 11 September 2026
AI agent governance means more than access control

Insights

AI agent governance means more than access control

Most AI-native PAM covers who gets access. Governance also has to judge what happens next.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 10 September 2026
Verification guidance exists. Turning it into a rule is still your job.

Insights

Verification guidance exists. Turning it into a rule is still your job.

A session analysis can tell you what to tighten. Turning that into a rule is still your job today.

Jungyeon Lee
Jungyeon LeeContent Marketer · 9 September 2026
Product update: what gets decided before a command runs

Product

Product update: what gets decided before a command runs

Choose how far commands are gated, confirm the server account during approval, and give automation credentials of its own.

Taeyeong Baek
Taeyeong BaekGTM Associate · 8 September 2026
Why AI governance pilots don't survive scale

Insights

Why AI governance pilots don't survive scale

The policy wasn't wrong. It just wasn't sized for agents your other teams stood up without telling you.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 7 September 2026
Out-of-band approval: why an agent's own channel can never be the one that approves

Insights

Out-of-band approval: why an agent's own channel can never be the one that approves

An AI agent that can approve a review on its own execution channel isn't gated at all. Why out-of-band approval separates the channel, not the person.

Marco Kwak
Marco KwakHead of GTM · 28 August 2026
HIPAA day-90 audit: what session expiry misses

Insights

HIPAA day-90 audit: what session expiry misses

A HIPAA day-90 audit should review grant history, not just a session's current expiry. Here is what to pull.

Marco Kwak
Marco KwakHead of GTM · 27 August 2026
Kubernetes access control governs access, not commands

Engineering

Kubernetes access control governs access, not commands

Access control checks whether you can reach a Kubernetes pod. It doesn't check whether the kubectl command you type should run.

Jungyeon Lee
Jungyeon LeeContent Marketer · 27 August 2026
Session-scoped sudo: bind OS-level root to a session, not a sudoers file

Engineering

Session-scoped sudo: bind OS-level root to a session, not a sudoers file

A sudoers rule only asks who's allowed to run a command. Sudo reaching Alpacon's command API asks a second question: is this command dangerous.

Jungyeon Lee
Jungyeon LeeContent Marketer · 26 August 2026
EU AI Act high-risk deadline moved to 2027. What didn't?

Insights

EU AI Act high-risk deadline moved to 2027. What didn't?

Article 50 stayed in force. Two new bans arrive in December. Only the classification clock moved—to 2027.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 26 August 2026
Two 2026 CVEs turned security tools into the attack path

Incident

Two 2026 CVEs turned security tools into the attack path

Microsoft Defender and Check Point SmartConsole both became attack paths in 2026, and one patch fell to a bypass a month later.

Marco Kwak
Marco KwakHead of GTM · 25 August 2026
Product update: what the audit log stops keeping—and stops missing

Product

Product update: what the audit log stops keeping—and stops missing

Command audit records are now safer, with more precise control over server access and approval requests.

Taeyeong Baek
Taeyeong BaekGTM Associate · 25 August 2026
SSO and MFA verify the login, not what the session does next

Incident

SSO and MFA verify the login, not what the session does next

A five-month Zimbra zero-day shows what happens after an attacker gets inside an authenticated session.

Jungyeon Lee
Jungyeon LeeContent Marketer · 24 August 2026
Credential theft: entitlements without purpose

Incident

Credential theft: entitlements without purpose

Storm-2949 turned one compromised identity into a cloud-wide breach—the gap a stolen credential's entitlements can never answer.

Marco Kwak
Marco KwakHead of GTM · 24 August 2026
Command filtering fails two ways: GTFOBins, and a shell that rewrites what you matched

Engineering

Command filtering fails two ways: GTFOBins, and a shell that rewrites what you matched

Measured across 1,709 real-world denylists, most don't fully block what they were written to block.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 20 August 2026
Offboarding access revocation: what a departing engineer leaves running

Insights

Offboarding access revocation: what a departing engineer leaves running

Disabling the account ends their login. It doesn't touch what they built, or tell you what they ran.

Marco Kwak
Marco KwakHead of GTM · 19 August 2026
AI agent execution control: what it is, and why AI-native PAM needs it

Insights

AI agent execution control: what it is, and why AI-native PAM needs it

Every agent constraint is either declared or enforced. Execution control is the enforced kind.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 18 August 2026
AgentForger: how one link forged a rogue AI agent with a borrowed employee session

Incident

AgentForger: how one link forged a rogue AI agent with a borrowed employee session

One crafted URL stood up an attacker-controlled AI agent inside the org. No stolen credential, no malware.

Jungyeon Lee
Jungyeon LeeContent Marketer · 17 August 2026
AI vendor risk assessment in 2026: the runtime questions your questionnaire is missing

Insights

AI vendor risk assessment in 2026: the runtime questions your questionnaire is missing

Your questionnaire covers who the vendor is and how they store data. It says nothing about what their agent executes on your servers once you authorize it.

Marco Kwak
Marco KwakHead of GTM · 14 August 2026
ISO 42001 Stage 1 checklist: the documentation to have ready before your audit

Insights

ISO 42001 Stage 1 checklist: the documentation to have ready before your audit

Stage 1 is a documentation and readiness review—the part of ISO 42001 you can actually prepare for in advance. Here's what an auditor expects on the table before Stage 2.

Marco Kwak
Marco KwakHead of GTM · 13 August 2026
Zero standing privilege by design: make it a property of the session, not the token

Engineering

Zero standing privilege by design: make it a property of the session, not the token

Most vendors make the credential ephemeral. We make the privilege itself session-bounded—so standing privilege can't accumulate, even while a session is live.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 13 August 2026
ISO 42001 audit checklist: the controls a policy document can't satisfy

Insights

ISO 42001 audit checklist: the controls a policy document can't satisfy

You can pass Stage 1 on paperwork. Stage 2 asks what your AI agents actually did in production—and three Annex A controls decide whether you have the answer.

Marco Kwak
Marco KwakHead of GTM · 12 August 2026
Alignment isn't enough: containing a misaligned agent's actions takes runtime execution control

Incident

Alignment isn't enough: containing a misaligned agent's actions takes runtime execution control

Model-level alignment lowers the odds an agent misbehaves. It can't drive the odds to zero once the agent is running on real infrastructure.

Jungyeon Lee
Jungyeon LeeContent Marketer · 12 August 2026
AI governance on paper vs. governance during the task

Insights

AI governance on paper vs. governance during the task

A near-miss doesn't change your policy. It changes the question leadership asks about it.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 11 August 2026
Product update: the judgment behind every approval

Product

Product update: the judgment behind every approval

See the reasoning behind each approval request, without bringing agent runs to a halt.

Taeyeong Baek
Taeyeong BaekGTM Associate · 11 August 2026
An autonomous AI agent breached Hugging Face—here's the kill chain, and where execution control bounds it

Incident

An autonomous AI agent breached Hugging Face—here's the kill chain, and where execution control bounds it

An autonomous AI agent ran a real end-to-end intrusion on Hugging Face's production infrastructure. Here is the kill chain, and an honest map of where a runtime execution-control layer bounds it.

Jungyeon Lee
Jungyeon LeeContent Marketer · 10 August 2026
LegacyHive: a Windows zero-day you can't patch your way out of

Incident

LegacyHive: a Windows zero-day you can't patch your way out of

No CVE, no vendor patch, and it works on fully-patched machines. When patch-and-wait isn't a control, what's left is the layer that governs privileged actions as they run.

Jungyeon Lee
Jungyeon LeeContent Marketer · 7 August 2026
AutoJack: one webpage, one MCP socket, host-level access

Insights

AutoJack: one webpage, one MCP socket, host-level access

Microsoft disclosed a three-stage exploit chain in AutoGen Studio that turns an AI browsing session into a remote code execution primitive. No single patch makes the underlying problem go away.

Jungyeon Lee
Jungyeon LeeContent Marketer · 28 July 2026
A command log won't tell you what your agent actually did.

Insights

A command log won't tell you what your agent actually did.

A command log tells you what ran. An audit has to answer who, why, and with whose approval—for humans and agents alike.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 21 July 2026
JIT access control is for humans. What should we have for AI agents?

Insights

JIT access control is for humans. What should we have for AI agents?

Just-in-time access governs whether an agent gets in, and when. It says nothing about what the agent does once inside.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 14 July 2026
Adding AI agents shouldn't mean adding more security exposure

Insights

Adding AI agents shouldn't mean adding more security exposure

Expanding your attack surface isn't the price of an AI agent. It's an artifact of how you wire it in – and it's avoidable.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 9 July 2026
The gate worked. The database was still dumped.

Incident

The gate worked. The database was still dumped.

Sysdig documented the first in-the-wild LLM agent intrusion: four pivots, every access gate cleared, database exfiltrated. The layer neither covers.

Eunyoung Jeong
Eunyoung JeongFounder & CEO · 24 June 2026
Work Sessions: the audit unit your CISO actually needs

Product

Work Sessions: the audit unit your CISO actually needs

Alpacon's Work Sessions turn fragmented audit timelines into one object – declared intent, approval gate, session-scoped sudo, unified record.

David Calvert
David CalvertDeveloper Advocate · 7 May 2026