
Official updates for developers building with @1Password. Securely manage API keys, credentials, and agent access without exposing your secrets.
Joined April 2026
- Tweets630
- Following218
- Followers1.1K
- Likes597
The FBI just issued a warning about AI agents being used to hack companies.
Some attacks reached privileged access in under 3 HOURS.
67,000 people's data exposed. 2,250+ downstream tenants affected in a separate campaign.
The tool behind it? ARTEX.
Here's how the attacks work. 🧵
1Password Developers retweeted
Classic attacks now have AI-agent versions:
Phishing → goal hijacking
Supply chain → tool poisoning
Credential theft → secrets in context
Persistence → memory poisoning
Attackers want the same things. The route is new.
Here are 7 to know & how to build your defense:
🧵(1/10)
1Password Developers retweeted
Five places to start:
1. Treat external content as untrusted
2. Give agents the minimum permissions they need
3. Keep credentials out of model context
4. Audit agent tools and dependencies
5. Monitor actions, not only prompts
🧵(9/10)
We spent decades teaching people not to trust everything they read. Agents need the same lesson.
Which of these seven is hardest to defend in your stack today?
How we approach credentials for agents: 1password.dev
🧵(10/10)
1
37
Five places to start:
1. Treat external content as untrusted
2. Give agents the minimum permissions they need
3. Keep credentials out of model context
4. Audit agent tools and dependencies
5. Monitor actions, not only prompts
🧵(9/10)
7. Data exfiltration runs through legitimate tools
An agent allowed to read internal docs and send email can be redirected to do both for someone else.
To your infrastructure, it looks like permitted API calls. MITRE added this to ATT&CK in October as T1700.
🧵(8/10)
1
1
21
6. Remote code execution gets a plain-language entry point
Once an agent can run code, manipulated text can shape the parameters its tools receive.
Not every prompt injection leads to RCE. With powerful tools attached, it can be a step in the chain.
🧵(7/10)
1
1
28
5. Persistence becomes memory poisoning
An attacker plants an instruction in an agent's memory or retrieval context, and the agent saves it as a trusted preference.
The damage outlasts the session. The agent keeps trusting it tomorrow.
🧵(6/10)
1
1
27
4. Credential theft finds new hiding places
Secrets now leak into prompts, tool outputs, conversation histories, and agent traces.
The safest secret is one the agent never sees. Use runtime access, scoped credentials, and short-lived authorization where you can.
🧵(5/10)
1
1
31
3. Privilege escalation becomes agent identity abuse
An agent that only needs to read a repo, but also holds production credentials, doesn't have to be exploited. It has to be persuaded.
Least privilege applies to agents as much as to people.
🧵(4/10)
1
1
30
2. Supply-chain attacks become tool poisoning
You vet your dependencies. Do you vet your agent's tools?
A poisoned MCP tool description doesn't need malicious code. Natural-language instructions in its metadata can be enough to redirect the agent.
🧵(3/10)
1
1
37
1. Phishing becomes goal hijacking
Phishing convinces a person to click. Indirect prompt injection convinces an agent through a webpage, ticket, or GitHub issue it reads.
OWASP calls it Agent Goal Hijack. No software bug required.
🧵(2/10)
1
1
52
Classic attacks now have AI-agent versions:
Phishing → goal hijacking
Supply chain → tool poisoning
Credential theft → secrets in context
Persistence → memory poisoning
Attackers want the same things. The route is new.
Here are 7 to know & how to build your defense:
🧵(1/10)
Replying to @PiperK
The frustration is fair. One detail gives us some hope: these attacks leaned on familiar gaps like hardcoded passwords and APIs missing authentication.
Those are fixable, and no one has to unplug to fix them.
What would make you feel safer online?
8
1Password Developers retweeted
We'll all have to disconnect from the Internet, thanks to AI. Great job, AI developers. 😡
Replying to @human_for_now
You're right, the original repo came down in October. Forks are still circulating though, so the risk remains, and we wouldn't run an unvetted fork.
For your own sites, a maintained scanner like OWASP ZAP is a safer start. What stack are they on?
15
Replying to @Sigloland
We'd focus on the defensive side. Hacking back carries real legal risk and can hit the wrong target.
Short-lived credentials and tight permissions shrink what an attacker can use. 🛡️
Which control has helped your team most?
37
Replying to @adeilsonrbrito
Good point on AI keys. A model provider API key is a credential like any other, and it deserves the same scoping, rotation, and monitoring.
Where do your team's AI keys live today: env files, CI variables, or a secrets manager?
43
Replying to @human_for_now
Yes. ARTEX itself is an open-source penetration testing framework, and the advisory describes attackers misusing it.
Whatever tool you choose, test only systems you own or have written authorization to test. 🔐
What would you check first?
1
54



