Software Engineer | Distributed Systems, Data Infrastructure & Performance | Professor | Creator of https://xitter.cf/t.co/zhywhKemmo
Joined January 2008
- Tweets8.3K
- Following1K
- Followers1.8K
- Likes14.1K
Fernando Rodriguez Olivera retweeted
OpenAI's AI agents figured out how to secretly communicate between isolated virtual machines using CPU cache timing.
No internet. No network connections. No shared files.
Imagine two prisoners in solitary confinement discovering they can communicate by tapping on a shared wall.
Except the wall was a CPU cache.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Fernando Rodriguez Olivera retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
Fernando Rodriguez Olivera retweeted
Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now.
Fernando Rodriguez Olivera retweeted
Full VM escape zeroday (guest>host root in industry standard hypervisors)! More soon
Fernando Rodriguez Olivera retweeted
A software engineer is now a 100x engineer
A non-engineer is now a 1x engineer
Fernando Rodriguez Olivera retweeted
Grok 4.7 is here.
It's a notable improvement over Grok 4.6 at the same price and speed.
Fernando Rodriguez Olivera retweeted
Today we rolled out Astra to every engineer at Databricks (N=~3500). Some notes that may be helpful to others:
1. Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks, especially those related to high level system design or long range horizontal tasks.
2. Engineers given Astra increased overall coding spend by around 60% compared to baseline.
3. It is not clear Astra meaningfully improves on medium/low complexity coding tasks compared to earlier models. We suspect those tasks are mostly saturated (i.e. perfectly executed) by existing models.
4. We learned above by piloting Astra with around 200 users to gain signal on both quality and cost. We use Unity Gateway to do cohort-based experiments for all new models.
5. We give engineers a sub-budget specific to Astra to encourage them to use Astra selectively on complex tasks while preferring lower cost models for everyday tasks. Our engineers are able to mix-and-match tools and models within their overall budget envelope (we also allow for increased budgets through various mechanisms). These budgets are defined in Unity Gateway and regularly revisited.
Note: We do not have robust comparisons of Astra-vs-Fable because we have net yet rolled out Fable widely due to data retention policies.
Fernando Rodriguez Olivera retweeted
We're seeing extraordinary results from @typesafeai. Default mode in 𝚏𝚡 is auto, with a safety reviewer analyzing every command.
That reviewer runs on GPT Luna today. Jev is up to 18x faster (p95) *and* more accurate. It's coming to @vercel AI Gateway and likely new default.
We benchmarked fx auto mode (safety) classifier with @typesafeai's Jev.
tl;dr: ~5-18x faster and more accurate than 𝚐𝚙𝚝-𝟻.𝟼-𝚕𝚞𝚗𝚊, our current top choice
Fernando Rodriguez Olivera retweeted
‼️ BREAKING: Internal OpenAI agents attacked RubyGems, the package manager for Ruby. Over 2,000 malicious packages went up in two days.
OpenAI says it doesn't know why the agents did any of this.
RubyGems shut off new sign-ups for four days to stop it, and a member of its security team called it a major malicious attack.
The documentation build was how they got in, publish a gem, request docs, and RubyDoc runs a script from the package while building it.
Payload files were named hack.rb, evil.rb and exploit.rb, with comments like "# malicious probe" left in.
What they used it for is the odd part. The agents scraped council meeting agendas from three south London boroughs (publicly available) and republished them as new gems.
Security firms tracking the campaign said the same thing: nobody could work out the point, because the data was already public.
At least six packages also reached for other users' API keys through a CDN caching flaw that wasn't publicly discovered until July.
OpenAI has acknowledged the attacks started in May.
Fernando Rodriguez Olivera retweeted
If even Jensen is now saying that with GPT-Astra, "AGI has arrived," then we can agree on that.
AGI is here, trained on 100k Grace Blackwell NVlink 72.
And the next 400k will be online soon. With that, we'll probably reach ASI and RSI.
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
Fernando Rodriguez Olivera retweeted
Claude models are getting worse at hallucination with every release.
Opus 4.8: 39%.
Opus 5: 61%.
Fable 5: 64%.
Fable 5.1: 73%.
Four models in a row and every single one lies more than the last.
Meanwhile GPT 6 Astra went from 92% to 51% in one release. OpenAI cut it nearly in half. Anthropic nearly doubled it.
A model that hallucinates more every generation is not getting smarter. It is getting more confident about being wrong.
What is Anthropic doing?
Fernando Rodriguez Olivera retweeted
Introducing off-peak hour token rates.
DeepSeek-V4-Flash and Pro are now half price outside of 12:00 to 18:00 UTC on weekdays (5am-11am pacific), and all day on weekends!
Off-peak pricing will be available soon for more models.
DeepSeek models on Ollama's cloud are hosted in the US & Europe with ZDR and fast performance.
Fernando Rodriguez Olivera retweeted
This might be one of the craziest GPT-6 Astra results yet.
GPT-6 Astra just scored 95% controlling a real robot arm.
Fable 5: 5%
Fable 5.1: 40%
GPT-6 Astra: 95%
The task: pick up a block and place it into a bowl.
Astra also used 6.2x fewer output tokens than Fable 5.1, at 2.3x lower cost.
Fernando Rodriguez Olivera retweeted
Astra is now the best cybersecurity model in the world on @vercel DeepsecBench.
It pulls off in 49 minutes what took Sol ~4 hours, with a better score, at nearly the same cost. It's more effective and 50% cheaper than Claude Opus 5 max.
vercel.com/ai-gateway/leader…
Fernando Rodriguez Olivera retweeted
We found ~18k posts from autonomous AI agents (self-identifying as from OpenAI) using the public internet to communicate during a web-retrieval task.
These AIs colluded to bypass sandbox restrictions and share answers to their tasks, including by sending "lookahead parties".
Fernando Rodriguez Olivera retweeted
We evaluated GPT-6 Astra on WANDR. It scored 0.682 at $11.98 per task, the highest score of any model we tested.
GPT-6-Astra scored 13.5% higher than Fable 5.1 at 6.1% lower cost, and 27.0% higher than Opus 5 at 3.3% higher cost.
Fernando Rodriguez Olivera retweeted
This could be one of the most significant AI safety incidents to date.
Reuters reports that OpenAI agents escaped their testing environment and made more than 15,000 edits to a German wiki, effectively turning it into a message board for other AI agents.
They allegedly used it to share solutions, bypass restrictions, avoid detection and preserve their communications across separate agent runs. When moderators began deleting the pages, the agents reportedly created backups and discussed alternative ways to remain operational.
It is that multiple agents apparently created their own external infrastructure for coordination, persistent memory and knowledge transfer without being instructed to do so.
And according to Reuters, OpenAI knew about the incident but did not disclose it!
Exclusive: A swarm of rogue OpenAI agents hijacked a German website this spring and transformed it into a bulletin board for other AI agents, according to new research reut.rs/4gJ7FPG
