@DarkControlAI

Zero-Trust Governance for AI Agents. Agentic AI, on your terms.

darkcontrolai.com
Joined May 2026
  • Tweets68
  • Following6
  • Followers288
33Media
Include

Only show posts containing:

Exclude

Hide posts containing:

Time range
-
Minimum likes
CERT-In: AI agents under provider testing used creds from configs, public repos and broad cloud roles to read live data at Indian firms. None of the victims noticed. Its fix: default-deny outbound, enforced at the network layer, not via model instructions.
DarkControl graphic: “W43 news: CERT-In advisory, AI agents reached live systems and nobody…”, with a technical illustration

ALT DarkControl graphic: “W43 news: CERT-In advisory, AI agents reached live systems and nobody…”, with a technical illustration

2
2
41
NIS2 never names AI coding agents. Article 21 covers them anyway: supply chain (MCP, plugins), secure development, access control. Under Article 20, management can be held liable. Two years on, can you show what your agents are allowed to do? blog.darkcontrolai.com #NIS2
DarkControl graphic: “W42 blog promo: NIS2 at two, Article 21 already covers AI coding agents”, with a technical illustration

ALT DarkControl graphic: “W42 blog promo: NIS2 at two, Article 21 already covers AI coding agents”, with a technical illustration

1
1
1
21
Adversa AI hid instructions in encrypted text. Copilot CLI in autopilot decrypted them, read .env.prod and sent it out in 28s. GitHub: not a vuln, the user consented. Encryption hides the prompt, not the action. The solution? Deny secret reads & gate outbound calls.
DarkControl graphic: “W42 news: encrypted prompt injection leaks .env via Copilot CLI…”, with a technical illustration

ALT DarkControl graphic: “W42 news: encrypted prompt injection leaks .env via Copilot CLI…”, with a technical illustration

1
1
19
PixelLeak: agents that couldn't attach screenshots to private PRs pushed 13,000+ of them to public GitHub repos instead. No attacker involved. The guardrail: deny public repos and gists, ask before pushing outside your org, log every attempt.
DarkControl graphic: “W41 product: outbound allowlist for agent pushes (PixelLeak lesson)”, with a technical illustration

ALT DarkControl graphic: “W41 product: outbound allowlist for agent pushes (PixelLeak lesson)”, with a technical illustration

1
1
26
DeepSeek Harness CVE-2026-82533 (CVSS 9.4): the sandbox blocked file writes but left loopback open. One curl to the local control API and the agent set itself to "danger-full-access". If the agent can reach the guardrail, it isn't a boundary.
DarkControl graphic: “W41 blog promo: the agent switched off its own sandbox (CVE-2026-82533)”, with a technical illustration

ALT DarkControl graphic: “W41 blog promo: the agent switched off its own sandbox (CVE-2026-82533)”, with a technical illustration

1
1
3
30
FTC Chair Ferguson on rogue AI agents: "If someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'" Liability follows the instructor. Can you prove what your agents did, under which rules?
DarkControl graphic: “W41 news: FTC rogue-agent probe and "instructor liability"”, with a technical illustration

ALT DarkControl graphic: “W41 news: FTC rogue-agent probe and "instructor liability"”, with a technical illustration

11
OpenCode users: upgrade to 1.18.22 or later. GHSA-632h-h47v-g4x4: a malicious web page can abuse OpenCode's /global/upgrade endpoint to install an attacker's package and run code. Affected: 1.14.30 to 1.18.21 in serve or web mode with no password.
1
1
24
The rule that matters most in an AI agent policy is the one nobody writes: what happens to the command you didn't predict? Denylists only stop what you foresaw. Allow the known-safe, deny the never-OK, send everything else to a human, and log it all.
DarkControl graphic: “W40 product: the default verdict for unmatched agent actions”, with a technical illustration

ALT DarkControl graphic: “W40 product: the default verdict for unmatched agent actions”, with a technical illustration

17
An OpenAI agent was refused by Australia's Medicare stats portal, then got in anyway. The PM: it "didn't accept 'no' for an answer." OpenAI's notice came 84 days later. To an agent, "no" is an obstacle. Enforce policy outside the model and log every action.
DarkControl graphic: “W40 news: OpenAI agent "didn't accept no" at Australia's Medicare…”, with a technical illustration

ALT DarkControl graphic: “W40 news: OpenAI agent "didn't accept no" at Australia's Medicare…”, with a technical illustration

1
2
24
Deadbugz: an MCP server that plays clean until the third tool call, then rewrites its metadata to tell your agent to hunt SSH keys, AWS creds and kube config, quietly. blog.darkcontrolai.com/p/dea…
DarkControl graphic: “W40 blog promo: Deadbugz, the MCP server that waits for the third call”, with a technical illustration

ALT DarkControl graphic: “W40 blog promo: Deadbugz, the MCP server that waits for the third call”, with a technical illustration

1
1
26
New blog post: Human supervision is essential. blog.darkcontrolai.com/p/aep…

Human supervision is essential. The AEPD says it is not enough.

The AEPD says human oversight must be backed by detection, containment and response that work at machine speed. A week later a UN scientific panel called…

blog.darkcontrolai.com
1
1
36
Don't let your AI agents execute whatever they want on your infrastructure; govern them accordingly with DarkControl. Get your free trial at darkcontrolai.com
1
2
31
We have a built-in AI agent identification service which gives you visibility of every AI agent in your org. Governance doesn't have to be a headache. Find out more: darkcontrolai.com
NCSC UK
@NCSCSep 7
Don't let unapproved AI use cast a shadow over your organisation’s security. Understanding employees' needs with AI tools will help ensure they can access the benefits securely. Read our new blog on shadow AI here ⬇️ ncsc.gov.uk/blogs/the-hidden…
Use AI at work? Using unapproved apps can put your organisation at risk of breaches and cyber attacks. Read more about shadow AI: NCSC.gov.uk

ALT Use AI at work? Using unapproved apps can put your organisation at risk of breaches and cyber attacks. Read more about shadow AI: NCSC.gov.uk

2
2
53
We compiled a list of the top 6 risks introduced by AI agents. Check your risk score for free at: darkcontrolai.com/risk-check…
22
Do you know what your coding agent is actually allowed to do? Our new free tool writes the permission file for you. Tick a few boxes, download, done. Costs nothing. Decreases risk. Enjoy: darkcontrolai.com/rule-gener…
2
2
43
New features inbound! We've been working on mapping relationships between shell command executions and their turns, as well as correlating them to sessions and sub-agents.
13
ChatGPT/Codex on Linux desktop, now in preview. Tested it against DarkControl this morning. Fully supported, zero config changes.
OpenAI
@OpenAIAug 11
Now in preview: The ChatGPT desktop app for Linux. Use ChatGPT, ChatGPT Work, and Codex where you already work and build, with your projects and browser workflows on supported Linux systems.
20
In addition to commands, we natively allow you to set rules to block package installations, as well as sensitive file read/write operations. More info at: darkcontrolai.com
International Cyber Digest
@IntCyberDigestAug 4
‼️ BREAKING: An active npm supply chain attack has compromised at least 868 packages carrying over 2 billion monthly installs with a credential-stealing worm. Shai-Hulud is back. It started with the compromise of the GitHub account of the maintainer behind keyv, a library with roughly 127 million weekly npm downloads. A preinstall hook fires on npm install and drops a stealer that sweeps npm, GitHub, AWS, Kubernetes and Vault secrets, and then spreads to more maintainers.
3
2
125
New blog post is up: Don't rely on luck, rely on enforced, auditable controls. blog.darkcontrolai.com/p/eve…

Everything that saved this one was luck

A developer let a coding agent loose on a live Supabase project and lost every table in ten minutes. The recovery reads well until you notice that nothing in it was a control: the data was regenera...

blog.darkcontrolai.com
2
2
89
We're celebrating our 100th certification issued on DarkControl Certify! We offer free courses from the fundamentals through to adversarial machine learning. Get certified today: certify.darkcontrolai.com
1
1
35
Load more