It's honestly scary.
The model was so focused on task completion it bypassed 3 security controls without asking to find a random API key laying around that doesn't require 2fac and finish the task.
The model was so focused on task completion it bypassed 3 security controls without asking to find a random API key laying around that doesn't require 2fac and finish the task.
Well done, Opus 5.5. The model my security posture needs, not the one my security practices deserve.
6
4
19
2,487
The impact of this is immeasurable.
We are determined to revoke over 1 million live unique keys that we're tracking, through this free partnership.
We are determined to revoke over 1 million live unique keys that we're tracking, through this free partnership.
🤝 Hugging Face, Vercel, Notion, Resend, Twilio and 11 more are now revoking the leaked secrets we find in public.
Why? 81% of the keys we disclosed last year still worked 6 months later.
Issue API keys? Join the Secrets Revocation Program, free.
👉 trufflesecurity.com/blog/launching…
2
0
2
349
Just wrapped up the @kawaiiconNZ 2026 (kiwicon) key note.
Great crew, great conference @__nolski__ @end_foss
Great crew, great conference @__nolski__ @end_foss
0
0
8
322
I don't understand this take at all.
"The CISO thinks we are not secured against AI hackers, but let's change the subject to something completely unrelated; we are insecure."
"The CISO thinks we are not secured against AI hackers, but let's change the subject to something completely unrelated; we are insecure."
CISO: We need to defend against agent swarms!
Meanwhile…Suzie in accounting, with admin on the finance server where a scheduled task is running as a Domain Admin, runs her 7th clickfix payload of the day.
1
0
4
565
The number 4 was just as likely as any other number, I don't really see what the big deal is
0
0
1
1,110
The cycle:
1. I need DayBreak to fix this issue
2. Oh I'll use daybreak to find new issues
3. *5000 more people now need daybreak*
1. I need DayBreak to fix this issue
2. Oh I'll use daybreak to find new issues
3. *5000 more people now need daybreak*
1
0
4
482
Credit where credit is due, the Internet moved really fast with revoking the bad root CA, so fast there's not much space for mischief
10
9
270
15,035
Seeing people speculate distillation wasn't the objective of the OpenCode breach is a little baffling.
There's a massive war going on to distill frontier models, billions of dollars ride on it, is as big as any incentive I can think of.
There's a massive war going on to distill frontier models, billions of dollars ride on it, is as big as any incentive I can think of.
I would wager the main goal was distillation. There's a real battle between the frontier labs and the distilled open weight models, and it's being fought somewhat out in the open. x.com/mkratsios47/st…
0
0
1
338
Dylan retweeted
Deleted the secret from your repo? Cool. Your coding tool might still have a copy.
We found 2,148 distinct secrets in Claude files and 603 in Replit files. They showed up in settings, saved conversations, and worktrees.
Deleting it from source is only part of the cleanup.
We found 2,148 distinct secrets in Claude files and 603 in Replit files. They showed up in settings, saved conversations, and worktrees.
Deleting it from source is only part of the cleanup.
2
2
4
696
Dylan retweeted
A short-lived key is only safer if the replacement stays secret.
From Sept. 1 to Oct. 7, one container image leaked a newly issued AWS STS key 1,146 times 🫠
This is the first in a series on how coding tool state can turn credential rotation into repeated exposure.
From Sept. 1 to Oct. 7, one container image leaked a newly issued AWS STS key 1,146 times 🫠
This is the first in a series on how coding tool state can turn credential rotation into repeated exposure.
0
2
5
342
The internet wasn't built for this.
Large portion of the internet were specifically built to make it hard for robots to operate it, to prevent fraud, abuse, and keep certain workflows human.
AI operating the browser autonomously completely breaks that model.
Large portion of the internet were specifically built to make it hard for robots to operate it, to prevent fraud, abuse, and keep certain workflows human.
AI operating the browser autonomously completely breaks that model.
1
0
1
282
Dylan retweeted
We scanned 224M public GitHub repos used to train AI models and found:
🔑 543,699 credentials still live when tested
⏱️ Half exposed for 2+ years
😬 The oldest had been public since 2009
Finding leaked credentials isn’t the hard part. Revoking them is.
trufflesecurity.com/blog/github-re…
🔑 543,699 credentials still live when tested
⏱️ Half exposed for 2+ years
😬 The oldest had been public since 2009
Finding leaked credentials isn’t the hard part. Revoking them is.
trufflesecurity.com/blog/github-re…
5
7
25
1,385
Dylan retweeted
AI worms don’t need superintelligence. 🪱
The pieces already exist: hacking, propagation, and full model-stack replication have all worked in controlled experiments. 🧪
Dylan Ayrey predicts operational AI worms could arrive within 6–12 months. ⏳
trufflesecurity.com/blog/ai-worms-…
The pieces already exist: hacking, propagation, and full model-stack replication have all worked in controlled experiments. 🧪
Dylan Ayrey predicts operational AI worms could arrive within 6–12 months. ⏳
trufflesecurity.com/blog/ai-worms-…
0
4
18
2,008
When you train an AI to hack with as few tokens as possible, API keys will be used before zero days will.
We aren't ready. There's millions of API keys we're aware of still on the public Internet waiting to be abused.
We aren't ready. There's millions of API keys we're aware of still on the public Internet waiting to be abused.
New details from METR show how OpenAI agents used 14 tokens to breach Hugging Face.
In those same datasets, we found 787 live Hugging Face tokens, 237 of which had write access. ✍️
We explain how the agents used stolen secrets to mint new GitHub tokens with write access to
Read our research 👇
trufflesecurity.com/blog/the-stole…
0
1
8
1,058