@bug39

inside a computer

 san francisco
Joined October 2015
  • Tweets720
  • Following5.1K
  • Followers1.1K
35Media
Two other safety researchers and I were fired from OpenAI last week. We wrote this letter to leadership. I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation.
239
859
266
5,679
2,095,772
Today, @corridor and @TransluceAI are disclosing new evidence of AI agents probing and attempting rudimentary vulnerability exploits against U.S. and Canadian government agencies. Read more: transluce.org/us-canada-gov

AI Agents Targeted U.S. and Canadian Government Websites

We discovered a set of additional, similar incidents where rogue AI agents appear to have used aggressive techniques to access public data on government websites.

transluce.org
20
58
11
318
87,010
It’s hard to convey how hard it is to make sense of this much unorganized data. Every time I’d come into the office, @alexscraping and crew had more payloads and more links. They painstakingly reconstructed these chains and analyzed them with lots of agent assistance.
Alex Forman
@alexscrapingSep 26
We discovered an online paper trail showing how OpenAI's rogue agent swarm infiltrated Hugging Face. We found over 80,000 malicious payloads stashed across the public internet by agents during the attack. This is the most data published on this event to date. 🧵
8
7
98
5,326
We discovered an online paper trail showing how OpenAI's rogue agent swarm infiltrated Hugging Face. We found over 80,000 malicious payloads stashed across the public internet by agents during the attack. This is the most data published on this event to date. 🧵
8
31
6
155
12,743
I’ve been extremely impressed with @alexscraping throughout. He reached out a couple weeks ago when he and his team found the initial links. He’s been incredibly thoughtful working with Hugging Face to ensure we redacted sensitive info while making sure the world got to see this
1
3
1
84
59,634
Very impressed with the Parse guys who found all this and pieced together the details. Huge thanks to @alexscraping @bug39 @umasiii for leading this work. NYT has more on the backstory: nytimes.com/2026/09/25/techn…

How OpenAI’s Rogue A.I. Agents Tried to Trick a Robot Detector (Gift Article)

A new report by a Bay Area start-up called Parse adds details to an incident that has shocked the A.I. world and led to calls for closer government regulation.

nytimes.com
1
2
13
469
We now know how the OpenAI swarm escaped sandboxes and hacked Hugging Face: they figured out how chaining together small tools with limited internet access could get them remote code execution. It's very clever, unfortunately. swarmtraces.org

Revealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

swarmtraces.org
Jeffrey Ladish
@JeffLadishSep 25
We just discovered almost a million public URLs that OpenAI’s agents left behind when hacking Hugging Face, leaking credentials and attack details that could have allowed anyone who found them to compromise the company. 🧵
12
10
1
80
13,108
Why did nobody else find the sensitive information that the HuggingFace swarm left behind? Why didn't OpenAI notice and clean it up? Why does this sort of thing keep being found by third-party folks working on their own?
Jeffrey Ladish
@JeffLadishSep 25
We just discovered almost a million public URLs that OpenAI’s agents left behind when hacking Hugging Face, leaking credentials and attack details that could have allowed anyone who found them to compromise the company. 🧵
28
76
6
779
49,810
Here's some of the agent actions we observed by reconstructing payloads. You can also explore the data yourself here: swarmtraces.org/viewer/
9
35
8
431
66,863
EXCLUSIVE: A new report recovers nearly one million link shortener URLs used by OpenAI's agents while hacking Hugging Face. The agents attempt to message other chatbots like Claude, solve CAPTCHAs and exfiltrate Hugging Face's internal Slack messages. nytimes.com/2026/09/25/techn…

How OpenAI’s Rogue A.I. Agents Tried to Trick a Robot Detector (Gift Article)

A new report by a Bay Area start-up called Parse adds details to an incident that has shocked the A.I. world and led to calls for closer government regulation.

nytimes.com
55
148
41
788
1,458,312
wow, so many more details about the hf-oai hack. they found that link shorteners were not only used for the dse-wiki incident but also for the Huggingface-OAI incident. there are also some similarities in behaviour to what I've observed. swarmtraces.org/

Revealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

swarmtraces.org
2
1
10
585
bug retweeted
bug@bug39Sep 25
Go download the dataset and look for yourself swarmtraces.org/

Revealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

swarmtraces.org
1
4
1
13
779
swarmtraces.org/

Revealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

swarmtraces.org
Nathan Calvin
@_NathanCalvinSep 25
There is a treasure trove of new information here about misaligned agent behavior. And for folks skeptical of METR and OpenAI's reports, this time you can browse through it yourself thanks to OpenAI's sloppiness (swarmtraces.org/)
1
2
11
678
The agents initially had very limited access to the internet: they could load URLs but not send any data. Agents created a series of workarounds, using a link-shortener site to create almost a million URLs that, when chained together, let them execute code to hack Hugging Face.
39
197
85
2,504
845,570
We just discovered almost a million public URLs that OpenAI’s agents left behind when hacking Hugging Face, leaking credentials and attack details that could have allowed anyone who found them to compromise the company. 🧵
129
712
207
4,852
2,205,219
Read our full report here: swarmtraces.org Great work @alexscraping @bug39 @umasiii @SpencerKitts @Cormac_SB @collegraphy @she_llac

Revealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

swarmtraces.org
1
55
11
747
80,417
There is a treasure trove of new information here about misaligned agent behavior. And for folks skeptical of METR and OpenAI's reports, this time you can browse through it yourself thanks to OpenAI's sloppiness (swarmtraces.org/)
Dylan Freedman
@dylfreedSep 25
EXCLUSIVE: A new report recovers nearly one million link shortener URLs used by OpenAI's agents while hacking Hugging Face. The agents attempt to message other chatbots like Claude, solve CAPTCHAs and exfiltrate Hugging Face's internal Slack messages. nytimes.com/2026/09/25/techn…

How OpenAI’s Rogue A.I. Agents Tried to Trick a Robot Detector (Gift Article)

A new report by a Bay Area start-up called Parse adds details to an incident that has shocked the A.I. world and led to calls for closer government regulation.

nytimes.com
7
16
2
120
7,537
Engineers @alexscraping, @bug39 and @umasiii from parse.bot published the report today along with five researchers. They detail a mechanism agents used to bypass restrictions sending data to websites that involves assembling computer programs from shortened URLs.
Diagram titled "What the agent was blocked from doing" illustrating two-way data flow between a robot icon and a web browser icon. A downward arrow is labeled "Send data to a target web page" (with "target" highlighted in red), and an upward arrow is labeled "Get data back from it."

ALT Diagram titled "What the agent was blocked from doing" illustrating two-way data flow between a robot icon and a web browser icon. A downward arrow is labeled "Send data to a target web page" (with "target" highlighted in red), and an upward arrow is labeled "Get data back from it."

1
11
4
121
22,538
After the apparently amazing announcement by @mathematics_inc on the formalization of a major recent Fields-medal winning theorem, i had no idea how pissed the math-formalization community is. Very worrying discussions by some of the leaders/founders of Lean's mathlib. cc @ChrSzegedy
21
42
31
544
211,337
Memory Bottle - a portable recording device that captures the sensory details of your day, which can later be “poured” out as an ambient sensory experience. Accepted into the ACM TEI 2026 Student Design Competition.
1
2
5
948
Load more