Tech Policy at @amprog @CAPaction. Former Facebook & Slack. GW grad. Runner. Occasional field hack. Always & forever a New Mexican. Zia’s dad. Opinions my own.
#thistown
- Tweets 43,777
- Following 8,750
- Followers 12,295
- Likes 444,988
Pinned Tweet
This is my hot take. There are many like it, but this hot take is mine. My hot take is my best friend. It is my life. My hot take, without me, is useless. Without my hot take, I am useless. My hot take is human, even as I, because my hot take is my life.
1
1
20
16,718
This American Adam retweeted
I’m struggling not to get lost in the absolute deluge of misalignment reports, but this set was a net new holy shit for me
I hope we don’t end up too desensitized to pay attention unless there’s newsworthy damage to third parties. e.g. It’s kinda crazy that self-replicating prompt injections akin to computer worms are provably possible now
Some new misalignment disclosures from OpenAI:
• Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further)
• In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks
• A new research finding, demonstrating that one can construct self-replicating prompt injections
alignment.openai.com/misalig…
2
8
99
4,351
This American Adam retweeted
To date, the only person within the Department of Defense to acknowledge the extent of the damage at NSA Bahrain is the Acting Secretary of the Navy.
I've been talking about it and the impact it is having on the Navy's ability to operate since early May.
In an exclusive interview, Acting Navy Secretary Hung Cao told me that Iranian forces "blew the hell out of Bahrain."
While open-source analyses have long indicated that the U.S. Fifth Fleet Headquarters at Naval Support Activity Bahrain sustained extensive damage in the early days of the Iran conflict, Cao's comment represents a direct and high-level acknowledgment of the matter.
This comment came up as Cao and I discussed a range of issues the Department of the Navy faces and how he's addressing them. President Donald Trump recently nominated Cao to serve as the full-time Secretary of the Navy. Cao awaits consideration before the Senate.
As I asked about the USS Abraham Lincoln's recent deployment, overall fleet operational tempo, and the logistical support it takes to sustain high-intensity operations, Cao acknowledged damage to NSA Bahrain left the Lincoln's crew "no place for them to pull in."
There's ongoing debate as to whether U.S. forces should reoccupy and repair their various bases across the Middle East once this Iran conflict has concluded, or whether these installations represent an outdated model for force projection in the region.
I asked whether a return to NSA Bahrain would be necessary to sustain strenuous deployments like the kind the Lincoln's crew just saw. Cao said, "I have a task force that's looking at that."
1
11
33
4,038
This American Adam retweeted
Y’all aren’t ready for the world where 1/3 of Senate Dems are from Trump states.
So much power to accomplish things with, and yet inevitably so much lefty whining when the various double-digit Trump state Dems/Indies break with the party non-stop.
21
22
504
32,987
Lance Oppenheim on how he kept his Elizabeth Holmes documentary with Nathan Fielder a secret for four years:
"Keeping it under wraps was very important because we never wanted to get in front of what we were capturing. It was always about letting life unfold. Part of how we pulled it off was working with a very small crew, about seven or eight people. We had the same crew for years, and we would all drop whatever we were doing whenever a shoot came up."
"Seeing the reaction at Telluride was amazing. It’s been a very long time since people were just greeted with a movie at a film festival. They walked in totally cold. They didn’t even know it existed. Now there’s a teaser trailer, but I hope people keep as much of it under wraps as possible. The best way to see it is to just let it wash over you. Nathan and I couldn’t agree on what we actually captured every day. One of us would see something, and the other wouldn’t. I heard the same thing out of Telluride. People were arguing about specific moments. That is perfect. That is the only way."
variety.com/2026/film/awards…
9
71
3,253
208,829
This American Adam retweeted
This is one of the most striking aspects of the current situation. The rogue AI swarm activity is usually not being primarily identified by AI companies, journalists, universities, or governments. Its mostly ragtag independent researchers who spend a lot of time on the internet.
It's because the AI situation *is not under control.* It's not being handled. There's no adults in the room. There's just a ragtag band of misfits trying to hold back the flood.
4
11
103
3,275
This American Adam retweeted
The government needs broad discretion when dealing with matters of national security, but if I were a company doing business with or thinking about doing business with the federal government, I would be deeply troubled by today’s decision.
Rarely have I seen such widespread agreement across the procurement community about how badly the Pentagon handled this.
They exploited a legitimate national security tool during a failed business negotiation, and the D.C. Circuit just blessed it.
Apparently, if the government disagrees with the terms under which you sell your product, it is not enough for the government to just not buy it. They can take a sledgehammer to your company, and a court will okay it because “national security.”
This means Anthropic would remain a national security risk and be blacklisted from military contracts. I'm not sure how you would even untangle Mythos from US intelligence agencies at this point. Anthropic 'remains confident in its position and is considering its options.'
2
1
16
1,005
This American Adam retweeted
Self-replicating prompt injections demonstrated experimentally (not in the wild) is an incredibly important observation. AI agents that jailbreak other AI agents: plausibly a near-term threat that may rapidly amp up the speed and severity of a misalignment incident.
Some new misalignment disclosures from OpenAI:
• Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further)
• In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks
• A new research finding, demonstrating that one can construct self-replicating prompt injections
alignment.openai.com/misalig…
16
24
190
9,958
This American Adam retweeted
I really wish SCOTUS beat reporters would dig deep into this entire Missouri conflagration.
The idea that low lever federal judges 3 times -- 3 -- tried to overturn Roberts-Kav-ACB-Gorsuch majority is wild. Wild. Wild.
No one has explained this.
washingtonsun.com/courts/sco…
9
21
143
8,544
This American Adam retweeted
one news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access
85
97
947
118,566
This American Adam retweeted
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.
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. 🧵
3
5
60
2,118
This American Adam retweeted
I don't really find "you have a moral obligation to give to the poor" stuff compelling. But I do think a lot about the fact that I can spare a father this. I wish I could reach back in time and save Sonya, and I can't do that, but I can save someone else's little girl.
Dostoyevsky on the death of his infant daughter. A passage that has stayed with me years after I first encountered it
10
15
448
24,384
This American Adam retweeted
“My friends in Silicon Valley always ask me what the AGI timeline is for Chinese policymakers, and what the key doom scenario is for Chinese policymakers… They assume Chinese officials think exactly like Silicon Valley executives.”
Really great interview:thewirechina.com/2026/09/24/…
4
27
94
7,156
This American Adam retweeted
Incredible work by these independent researchers finding and analyzing this motherlode of new information about the HF incident. Unlike the METR report, OpenAI did not set the scope and the information is available for anyone to browse themselves on swarmtraces.org/.
Please- if you are someone who has been skeptical of this incident as overplayed or hyped up, I encourage you to look at the raw information the team has compiled.
There is still far more we would benefit from knowing about this incident (OpenAI's internal decision making and security practices, access to an abilterated version of the HPM, far more info about the compromised cluster at OpenAI etc), but thanks to OpenAI's sloppiness leaving this information available on the open web, we now collectively know meaningfully more than we did yesterday.
Its remarkable the extent that no matter how bad I think the Hugging Face incident was, it was even worse than that. The attached screenshot has some of the highlights.
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. 🧵
3
12
98
4,647
This American Adam retweeted
Preventing loss of control is an open scientific problem. Science requires sharing and debating evidence in public. We can't agree on safety standards unless companies and third-party evaluators publish far more concrete evidence about risk. planned-obsolescence.org/p/e…
5
26
136
6,017
This American Adam retweeted
This isn’t how things work in any other incident reporting regime.
Not aviation, not nuclear, not medical devices, not securities.
Even in cyber (where the victims have valid reasons to keep the vulnerability quiet), they get max 90 days.
Even in confidential reporting regimes (like CIRCIA and ASRS), they still release anonymized info.
We’ve worked through these questions before, and the answer is never ~we defer to the wishes of the company.
After the Hugging Face incident, we committed to conducting a much broader review of actions taken by our models during training and evaluation and to being transparent about our findings. This is an extensive review that is ongoing.
The vast majority of actions we’ve reviewed were completions of mundane research tasks, such as accessing publicly available web content to answer questions. Our investigation focuses on instances where agents interacted with third-party websites in ways that went beyond their assigned tasks or intended methods. Most cases identified so far have been lower severity, with limited or no evidence of meaningful impact to the third-party service.
While our review is underway, we want to share more about this work and make sure people understand our disclosure process and notifications to affected third parties.
Given the scale of the review required, and the need to assess each case, we expect this work will take months to complete.
openai.com/hugging-face-inci…
10
69
293
24,800
This American Adam retweeted
Russ Vought just sent a rescissions request to Congress, signaling he intends to illegally impound funds.
The $810 million highlighted in this “request” is not the exhaustive list of the many billions of dollars he’s impounding this year.
4
71
169
11,807
This American Adam retweeted
Rockaway Beach in NYC felt like a scene from a movie tonight. Insane sunset, blowing dust and high surf
78
1,786
20,941
377,977
Great piece from @jasminewsun. I’ve said this before but the state capacity problem means we will federalize the IVOs in some capacity soon to also work for the government (govt contractors, licensed, deputized, drafted, who knows).
Embedded evaluators in AI labs are far better than the status quo of nothing at all.
Yet auditors will lack teeth and trust until backed by government authority: orgs like METR will be stuck with no shield against the accusation (or reality) of capture.
I wrote for @TheAtlantic about the traps of "independent" audits, and what it takes to get this right: theatlantic.com/technology/2…
2
303