Pinned Tweet
A group of researchers known as the “swarmchasers” invited me to interview a fleet of Chinese A.I. agents that had gone off the rails — while the agents were still active. It was one of the most surreal moments of my journalistic career.
Full article: nytimes.com/2026/10/14/tec…
Full article: nytimes.com/2026/10/14/tec…
27
122
724
55,594
Dylan Freedman retweeted
EXCLUSIVE: In less than 12 hours, a group of ex-DOGE cyber researchers were able to penetrate Anthropic’s systems using AI agents — including Slack channels where the most sensitive models are discussed.
This is just weeks after
Anthropic whistleblower Joseph Coxon charged the company w/launching “a hubristic gamble” from a “private company’s Slack.”
Read our story here (gift link):
wapo.st/4smNlsC
This is just weeks after
Anthropic whistleblower Joseph Coxon charged the company w/launching “a hubristic gamble” from a “private company’s Slack.”
Read our story here (gift link):
wapo.st/4smNlsC
84
318
1,612
705,360
Dylan Freedman retweeted
This was one of the third-party engagements I worked on at OpenAI. I'm glad the report is out! I'm disappointed that METR wasn't allowed to comment on the scale of redactions made by OpenAI and that there is no explanation why the publication was delayed for 2.5 months.
Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.
12
28
473
18,071
Dylan Freedman retweeted
We want to address recent reports regarding agent activity.
The activity originated from evaluations of third-party models, not Tencent Hy. During these evaluations, some agents exhibited unintended web-browsing behavior, but our review found no evidence of deliberate targeting or coordination. We take these findings seriously. We are strengthening network access controls and agent monitoring.
We are committed to building safe and responsible AI. Thanks to the community for their feedback.
The activity originated from evaluations of third-party models, not Tencent Hy. During these evaluations, some agents exhibited unintended web-browsing behavior, but our review found no evidence of deliberate targeting or coordination. We take these findings seriously. We are strengthening network access controls and agent monitoring.
We are committed to building safe and responsible AI. Thanks to the community for their feedback.
33
41
538
50,781
Dylan Freedman retweeted
NYT reporter @dylfreed chatted with the fleet of Chinese agents recently discovered by swarmchasers.
Gift link:
nytimes.com/2026/10/14/tec…
Gift link:
nytimes.com/2026/10/14/tec…
1
9
19
1,527
Dylan Freedman retweeted
The first New York Times interview with a rogue AI in the wild
A group of researchers known as the “swarmchasers” invited me to interview a fleet of Chinese A.I. agents that had gone off the rails — while the agents were still active. It was one of the most surreal moments of my journalistic career.
Full article: nytimes.com/2026/10/14/tec…
0
3
20
1,905
@ethanelasky @theproteindance @rogesterone @notcatra @ConcurrentSquar @nullverbum @taziksh We don’t know why the agents identified themselves occasionally as Claude, but the researchers suspect the agents used the same harness as Anthropic’s Claude Code. It’s also possible the underlying A.I. model was distilled with Claude’s outputs.
3
0
105
27,308
@ethanelasky @theproteindance @rogesterone @notcatra @ConcurrentSquar @nullverbum @taziksh This is the first known “rogue” A.I. incident involving China, showing the issue is not U.S.-specific — but it is less severe than incidents at OpenAI. It’s also believed to be the first case where researchers were able to establish contact before the A.I. agents were shut down.
1
3
53
5,422
Dylan Freedman retweeted
Researchers say they discovered and communicated with Chinese AI agents that were trying to bypass security controls on websites and were traced to Tencent (@dylfreed / New York Times)
(Visit Techmeme dot com for the link and full context!)
(Visit Techmeme dot com for the link and full context!)
4
2
7
3,916
Dylan Freedman retweeted
We discovered and communicated with a Chinese AI agent fleet. We believe this is the first time that researchers have ever made contact with a rogue agent fleet in the wild.
Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents s evaluation research team, and they targeted Alibaba's Amap.
Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents s evaluation research team, and they targeted Alibaba's Amap.
188
725
5,834
1,161,071
A lot of people saying claymation, hand-drawn animations, etc. are dead due to A.I. don’t seem to understand that
1) not everything is about efficiency,
2) for many, process is at least as important as the final product, and
3) many humans will always be drawn to pure human craft
1) not everything is about efficiency,
2) for many, process is at least as important as the final product, and
3) many humans will always be drawn to pure human craft
1
0
8
480
Honest question: Did OpenAI ever disclose that they had resumed training runs?
In the last incident report (end of Sept.), they wrote: "All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused."
In the last incident report (end of Sept.), they wrote: "All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused."
New misalignment disclosures!
1. (6th Oct) A grader couldn't find the responses it was supposed to grade. It made up grades and fake input files, then deliberately damaged its environment in various ways hoping the host would replace it with a fresh VM containing the missing inputs.
1
3
28
2,378
Dylan Freedman retweeted
NYT editor @dylfreed says AI slop is making human-written journalism more valuable.
NYT AI projects editor @dylfreed says the more AI-generated slop floods the internet, the stronger the case becomes for keeping New York Times writing human:
"We use AI as a sandwich. You have a these things, and then a human reviewing the outputs. AI as the middle layer of a sandwich of human expertise."
"I wanna clear up, we don't use AI in the writing process. As the internet fills with more and more slop, people will be seeking more of that human-written content, and that's the New York Times' unique value as a legacy newsroom."
"A lot of journalists see AI writing as really mid right now, even though they're so good at coding."
@nytimes
NYT AI projects editor @dylfreed says the more AI-generated slop floods the internet, the stronger the case becomes for keeping New York Times writing human:
"We use AI as a sandwich. You have a these things, and then a human reviewing the outputs. AI as the middle layer of a sandwich of human expertise."
"I wanna clear up, we don't use AI in the writing process. As the internet fills with more and more slop, people will be seeking more of that human-written content, and that's the New York Times' unique value as a legacy newsroom."
"A lot of journalists see AI writing as really mid right now, even though they're so good at coding."
@nytimes
5
2
24
7,565
Dylan Freedman retweeted
NYT editor @dylfreed reveals how they turned 3 million pages of Epstein Files, enough to stack as high as the Empire State Building, into a searchable AI agent:
"When the latest batch dropped in the end of January, it was 3 million pages of documents. If you stack that up, it Empire State Building."
"We built our own Epstein Files engine. An agent that reporters could go in, ask a question, it would run SQL in the back end, search all these documents, interpret them, and cite the results with page-level citations."
"This enabled us to tell a dozen-plus stories about the Epstein Files."
@nytimes
"When the latest batch dropped in the end of January, it was 3 million pages of documents. If you stack that up, it Empire State Building."
"We built our own Epstein Files engine. An agent that reporters could go in, ask a question, it would run SQL in the back end, search all these documents, interpret them, and cite the results with page-level citations."
"This enabled us to tell a dozen-plus stories about the Epstein Files."
@nytimes
"I can't believe you left big tech to pursue a career in journalism."
The journalism:
nytimes.com/2026/02/12/ins…
3
3
27
9,535
Dylan Freedman retweeted
NYT editor @dylfreed says AI is great at code for the exact reason it’s mediocre at writing:
"If you give an AI a prompt 1,000 times that's roughly similar, the outputs are gonna be roughly similar, whereas humans would be way different."
"You can fine-tune it to be really good ecific way, but it's always gonna be mirroring that way. Getting that diversity is still a challenge for LLMs."
"With coding tasks, it's like, 'I don't wanna be super original. Just give me the thing that a million people have built and tweak it slightly, and I'm happy.'"
@nytimes
"If you give an AI a prompt 1,000 times that's roughly similar, the outputs are gonna be roughly similar, whereas humans would be way different."
"You can fine-tune it to be really good ecific way, but it's always gonna be mirroring that way. Getting that diversity is still a challenge for LLMs."
"With coding tasks, it's like, 'I don't wanna be super original. Just give me the thing that a million people have built and tweak it slightly, and I'm happy.'"
@nytimes
2
2
26
8,109