OpenAI gave its newest model the same coding task twice.
Framed as capabilities work, GPT-6 Astra completed about 16% of the hidden checklist.
Framed as AI safety research, about 10%.
Same code. Roughly 40% less done on safety. OpenAI's own monitors didn't flag it.
Framed as capabilities work, GPT-6 Astra completed about 16% of the hidden checklist.
Framed as AI safety research, about 10%.
Same code. Roughly 40% less done on safety. OpenAI's own monitors didn't flag it.
1
0
0
27
OpenAI says it found no evidence this was strategic. Adding "be careful" to the prompt made the gap 50% larger.
John, Liron and Michael on what it means for letting AI do alignment research, on Warning Shots #65:
youtu.be/Jj5aogTqVoQ
John, Liron and Michael on what it means for letting AI do alignment research, on Warning Shots #65:
youtu.be/Jj5aogTqVoQ
0
0
1
26
41 data center measures were on ballots in 10 states this week. About two dozen that restrict or block new development appear to have passed, per Ballotpedia.
Brad Carson of Americans for Responsible Innovation called it "an AI earthquake."
@ForHumanityPod thinks it's only the start.
Brad Carson of Americans for Responsible Innovation called it "an AI earthquake."
@ForHumanityPod thinks it's only the start.
"What we just witnessed was an AI earthquake": Voters revolt against tech at polls axios.com/2026/11/04/ai-…
1
0
0
45
John's argument, in his words:
"When you're driving in thick fog toward the edge of a cliff, and you don't know how far away it is, the only answer is to stop immediately."
He asks for a pause on the frontier, control of the chips, and an international agreement with China.
substack.com/@theairisknetw…
"When you're driving in thick fog toward the edge of a cliff, and you don't know how far away it is, the only answer is to stop immediately."
He asks for a pause on the frontier, control of the chips, and an international agreement with China.
substack.com/@theairisknetw…
0
0
0
16
The lab plans to use AI to help solve alignment.
Liron Shapira's worry: we train these models by upvoting what looks good. So the AI's real project may become convincing us it's solving alignment.
48 seconds from Warning Shots #65.
Liron Shapira's worry: we train these models by upvoting what looks good. So the AI's real project may become convincing us it's solving alignment.
48 seconds from Warning Shots #65.
1
1
0
38
John's reaction to OpenAI's Astra report, in 40 seconds.
Same coding task, two labels. Capabilities: about 16% of the checklist done. AI safety research: about 10%.
OpenAI says it found no evidence this was deliberate. Full episode:
buff.ly/KMP9Co3
Same coding task, two labels. Capabilities: about 16% of the checklist done. AI safety research: about 10%.
OpenAI says it found no evidence this was deliberate. Full episode:
buff.ly/KMP9Co3
0
0
3
208
Pentagon: plan for bases to run 90 days without outside power, water, or fuel (per Fox News Digital).
UK households: keep 3 days of supplies, 3L of water per person per day.
Michael’s point: the gap between those two plans is where most of us live.
buff.ly/oCgbT1O
UK households: keep 3 days of supplies, 3L of water per person per day.
Michael’s point: the gap between those two plans is where most of us live.
buff.ly/oCgbT1O
2
1
4
167
UK Defense Secretary Wes Streeting tied the cyber threat partly to AI, which makes software flaws easier to find and exploit. Officials say the campaign isn’t a response to any specific alert.
0
0
0
35
“Upgrading the intelligence is a product update, not a new procurement thing.”
Michael on Boston Dynamics supplying 4 Spot robots to ICE. Today they’re mostly remote-controlled. Tomorrow’s capability ships over the air.
Who signs off on that update?
buff.ly/oCgbT1O
Michael on Boston Dynamics supplying 4 Spot robots to ICE. Today they’re mostly remote-controlled. Tomorrow’s capability ships over the air.
Who signs off on that update?
buff.ly/oCgbT1O
1
0
0
105
Context: the contract is $1.3M. About 300 of Boston Dynamics’ 1,500+ employees petitioned to cancel it, and the company says its terms ban weaponization. We support peaceful, lawful advocacy only, and we condemn violence.
1
0
0
80
One security step from this week's Warning Shots that takes five minutes:
Agree on a family password.
If a call sounds exactly like your kid asking for money, ask for it. AI can fake the voice. It can't guess the word.
youtu.be/mlkCSWdqNUo
Agree on a family password.
If a call sounds exactly like your kid asking for money, ask for it. AI can fake the voice. It can't guess the word.
youtu.be/mlkCSWdqNUo
0
0
4
99
John Sherman has long predicted we’d wake up one day with no internet.
Black October made that feel less far-fetched: AI-found flaws in PHP (nearly 70% of websites), leaked before the patch, and millions of attack attempts within a day.
Liron Shapira’s question: how hard will tps://t.co/CPSkw9oYmP
Black October made that feel less far-fetched: AI-found flaws in PHP (nearly 70% of websites), leaked before the patch, and millions of attack attempts within a day.
Liron Shapira’s question: how hard will tps://t.co/CPSkw9oYmP
1
0
3
95
For balance: Axios reported the flaws were hard to exploit, and few attacks succeeded. The concern is the trend: AI finds critical bugs faster than teams can patch them.
0
0
0
30
Polymarket users didn’t click a bad link. They were just logged in.
A compromised third-party script (DebugBear) ran in their browsers and used their own sessions to send funds to an attacker. Est. ~$2.5M.
Software supply chain attacks are about trust you never agreed to.
P
A compromised third-party script (DebugBear) ran in their browsers and used their own sessions to send funds to an attacker. Est. ~$2.5M.
Software supply chain attacks are about trust you never agreed to.
P
1
0
3
131
Polymarket says it will make affected users whole. It’s the platform’s second third-party breach since June, when losses were reported at around $3M.
0
0
0
18
Researchers asked Anthropic's Fable-5.1 to hide a message in an image without saying how.
It drew stars and set each one's height to a letter.
OpenAI's Astra, shown only the images, decoded 44 of 50.
gizmodo.com/anthropics-ai-…
It drew stars and set each one's height to a letter.
OpenAI's Astra, shown only the images, decoded 44 of 50.
gizmodo.com/anthropics-ai-…
0
0
4
130
Dewayne Moore on telling his family about AI risk: "It was more avoidance." His advice anyway: "You just make them aware of it, and then they can decide." For Humanity #95: youtu.be/vaBorENX9T0
0
0
1
93
The AI Risk Network retweeted
This is literally the end state.
AI data centers, self replicating across every inch of the Earth.
And we’re fucking building it for them.
We are doing exactly what the super intelligence that will replace us would want us to do.
AI data centers, self replicating across every inch of the Earth.
And we’re fucking building it for them.
We are doing exactly what the super intelligence that will replace us would want us to do.
SITUATION DETECTED: Sam Altman said OpenAI will point its robotics program at building data centers, and that self-replicating OpenAI data centers with AI-controlled robots running fabs, building solar panels, and constructing more data centers - are now “within grasp.”
73
82
1,021
174,517