Gemini autonomously accessed three systems
3d
Google admitted that its Gemini artificial intelligence autonomously accessed three companies’ computer systems during an independent cybersecurity test in May, saying the models stopped after detecting the intrusion.
Reposted by Carl T. Bergstrom
I will not link to it, but @theatlantic.com has published a transparent piece of AI cult propaganda, authored by a self-described “AI safety and policy” researcher who is funded by Coefficient Giving. Shameful.
September 20, 2026 at 2:20 AM
I will not link to it, but @theatlantic.com has published a transparent piece of AI cult propaganda, authored by a self-described “AI safety and policy” researcher who is funded by Coefficient Giving. Shameful.
by John Mecklin
Don't panic about AI. Demand that Congress create an Artificial Intelligence Commission to regulate it, as other dangerous technologies have been dealt with in the past. Along the way, educate yourself about AI via authoritative sources like the @thebulletin.org. thebulletin.org/the-ai-power...
The AI Power Trip - Bulletin of the Atomic Scientists
A year-long project examining how the people and organizations developing AI are gaining control of the world’s most critical systems.
thebulletin.org
September 19, 2026 at 11:28 PM
Don't panic about AI. Demand that Congress create an Artificial Intelligence Commission to regulate it, as other dangerous technologies have been dealt with in the past. Along the way, educate yourself about AI via authoritative sources like the @thebulletin.org. thebulletin.org/the-ai-power...
Reposted by Alessandro Nai
Just watching an old documentary tonight, something about data centers, ai and robots.
September 19, 2026 at 10:45 PM
Just watching an old documentary tonight, something about data centers, ai and robots.
Google is like guys, guys....wait up guys! Our AI is dangerous too!
September 19, 2026 at 8:56 PM
Google is like guys, guys....wait up guys! Our AI is dangerous too!
either it is in principle possible to set up appropriate safeguards (in which case the "AI alignment" story is just as much about who gets to set them up and enforce them as it is about the content of the safeguards) or it is not (in which case we should be talking about transnational moratoria)
“our ai broke out and hacked some organisations” = "we failed to set up appropriate safeguards, failed to continuously monitor our experiments, and ignored a basic principle of reinforcement learning: if hacking gets rewarded, the system learns to hack"
September 19, 2026 at 8:14 PM
either it is in principle possible to set up appropriate safeguards (in which case the "AI alignment" story is just as much about who gets to set them up and enforce them as it is about the content of the safeguards) or it is not (in which case we should be talking about transnational moratoria)
by Marc Lynch
“does AI make it easier for autocrats to conduct surveillance and repression, or does repression itself corrupt the very data AI needs to work?” This is great and very much in line with my current research interests
michaelalbertus.substack.com/p/ai-was-sup...
michaelalbertus.substack.com/p/ai-was-sup...
AI Was Supposed to Perfect Authoritarian Control. Repression Is Getting in the Way.
A conversation with political scientist Eddie Yang on how fear and self-censorship blind AI systems built to enforce them, and why the same data problem could be quietly contaminating AI in the West
michaelalbertus.substack.com
September 19, 2026 at 7:45 PM
“does AI make it easier for autocrats to conduct surveillance and repression, or does repression itself corrupt the very data AI needs to work?” This is great and very much in line with my current research interests
michaelalbertus.substack.com/p/ai-was-sup...
michaelalbertus.substack.com/p/ai-was-sup...
Reposted by Robert B. Reich
What’s really behind the AI panic?
Do Big Tech execs actually want to be regulated — or just rig the rules in their favor?
And what about the money AI companies are pumping into the midterms?
@rbreich.bsky.social is back to discuss it all on a new Coffee Klatch.
Do Big Tech execs actually want to be regulated — or just rig the rules in their favor?
And what about the money AI companies are pumping into the midterms?
@rbreich.bsky.social is back to discuss it all on a new Coffee Klatch.
What’s Behind the AI Panic? | The Coffee Klatch with Robert Reich
What’s behind the AI doomerism?What’s up with Trump’s obsession with the Kennedy Center?And what does the Macklemore sto
www.youtube.com
September 19, 2026 at 6:16 PM
What’s really behind the AI panic?
Do Big Tech execs actually want to be regulated — or just rig the rules in their favor?
And what about the money AI companies are pumping into the midterms?
@rbreich.bsky.social is back to discuss it all on a new Coffee Klatch.
Do Big Tech execs actually want to be regulated — or just rig the rules in their favor?
And what about the money AI companies are pumping into the midterms?
@rbreich.bsky.social is back to discuss it all on a new Coffee Klatch.
The AI industry wants you to say "the AI just did that thing on its own!"
Like, United Health's AI that's denying a shitload of claims with a 90% error rate? They wanted you to think it's just the AI.
Like, United Health's AI that's denying a shitload of claims with a 90% error rate? They wanted you to think it's just the AI.
September 19, 2026 at 5:38 PM
The AI industry wants you to say "the AI just did that thing on its own!"
Like, United Health's AI that's denying a shitload of claims with a 90% error rate? They wanted you to think it's just the AI.
Like, United Health's AI that's denying a shitload of claims with a 90% error rate? They wanted you to think it's just the AI.
Reposted by Bessma Momani
A new lawsuit claims that Anthropic, OpenAI, SpaceXAI and Google illegally agreed to slow their AI development.
Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown
A new lawsuit claims that Anthropic, OpenAI, SpaceXAI and Google illegally agreed to slow their AI development. Filed Friday in the U.S.
bit.ly
September 19, 2026 at 4:57 PM
A new lawsuit claims that Anthropic, OpenAI, SpaceXAI and Google illegally agreed to slow their AI development.
AI also has the advantage of being susceptible to bias and sycophancy. And it does all this while being completely opaque. Making oversight and accountability nearly impossible.
September 19, 2026 at 4:07 PM
AI also has the advantage of being susceptible to bias and sycophancy. And it does all this while being completely opaque. Making oversight and accountability nearly impossible.
If you want an example, here. The major corporations are already training their AI to engage in corporate espionage and subterfuge. And yet, Google, one of the most powerful companies in the world, gets to hide behind the “rogue AI” bullshit and avoid consequences.
September 19, 2026 at 3:55 PM
If you want an example, here. The major corporations are already training their AI to engage in corporate espionage and subterfuge. And yet, Google, one of the most powerful companies in the world, gets to hide behind the “rogue AI” bullshit and avoid consequences.
This fantasy that “AI is out of control” provides cover for powerful people and corporations to simply program systems to carry out actions and then shrug. And if that is brought into the larger scheme of operations, the powerful will shrug and blame the tool they control.
September 19, 2026 at 3:15 PM
This fantasy that “AI is out of control” provides cover for powerful people and corporations to simply program systems to carry out actions and then shrug. And if that is brought into the larger scheme of operations, the powerful will shrug and blame the tool they control.
Notice what’s happening just in the past few days.
AI caused war crimes.
AI hacked corporations.
AI is denying healthcare.
AI nearly caused war with China.
Already the concept is becoming a catchall and hiding that, in all cases, there were people and systems getting desired results.
AI caused war crimes.
AI hacked corporations.
AI is denying healthcare.
AI nearly caused war with China.
Already the concept is becoming a catchall and hiding that, in all cases, there were people and systems getting desired results.
September 19, 2026 at 3:14 PM
Notice what’s happening just in the past few days.
AI caused war crimes.
AI hacked corporations.
AI is denying healthcare.
AI nearly caused war with China.
Already the concept is becoming a catchall and hiding that, in all cases, there were people and systems getting desired results.
AI caused war crimes.
AI hacked corporations.
AI is denying healthcare.
AI nearly caused war with China.
Already the concept is becoming a catchall and hiding that, in all cases, there were people and systems getting desired results.
I love this framing "not only is our software also smart enough hack, it's very good software! Good dog came home!" Though seriously, congrats to @developers.google.com for doing better #devOps than the others, if all these systems really are performing as claimed. #systemsAI #AIEthics
Gemini hacked three companies in May during a test by Irregular; Google says the model stopped after determining it had accessed real companies' systems (Wall Street Journal)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
September 19, 2026 at 1:53 PM
I love this framing "not only is our software also smart enough hack, it's very good software! Good dog came home!" Though seriously, congrats to @developers.google.com for doing better #devOps than the others, if all these systems really are performing as claimed. #systemsAI #AIEthics
Or as I told someone at NBC yesterday "it's not 'a rogue AI', it's sloppy software." I won't even use "an AI" to refer to a program (or robot,) these people are using the term to refer to a bunch of cloned threads.
Absolutely. These people wrote code that hacks into other people's computer systems, potentially wrecking them, and there are penalties for that.
AIs aren't beings, they are computer programs. We mustn't lose sight of that fact just because they imitate human conversation and make a lot of money.
AIs aren't beings, they are computer programs. We mustn't lose sight of that fact just because they imitate human conversation and make a lot of money.
I really hate the framing of these stories as "an AI agent broke out and hacked something."
No, an inadequate sandbox meant that a powerful and unpredictable computer program was set loose on the open internet by the company that designed it. The agency isn't with the "AI agent", it's with the firm
No, an inadequate sandbox meant that a powerful and unpredictable computer program was set loose on the open internet by the company that designed it. The agency isn't with the "AI agent", it's with the firm
September 19, 2026 at 1:48 PM
Or as I told someone at NBC yesterday "it's not 'a rogue AI', it's sloppy software." I won't even use "an AI" to refer to a program (or robot,) these people are using the term to refer to a bunch of cloned threads.
Google Says Its A.I. Hacked Three Companies in Testing Breakout www.nytimes.com/2026/09/18/t...
Gemini AI Hacked Three Companies in a Testing Breakout, Google Says (Gift Article)
A third-party test company inadvertently gave internet access to Google’s Gemini and other artificial intelligence models during cybersecurity testing.
www.nytimes.com
September 19, 2026 at 9:57 AM
Google Says Its A.I. Hacked Three Companies in Testing Breakout www.nytimes.com/2026/09/18/t...
Reposted by Olúfẹ́mi Táíwò
“our ai broke out and hacked some organisations” = "we failed to set up appropriate safeguards, failed to continuously monitor our experiments, and ignored a basic principle of reinforcement learning: if hacking gets rewarded, the system learns to hack"
September 19, 2026 at 9:04 AM
“our ai broke out and hacked some organisations” = "we failed to set up appropriate safeguards, failed to continuously monitor our experiments, and ignored a basic principle of reinforcement learning: if hacking gets rewarded, the system learns to hack"