Jens Ernstberger (@0xSerious) header
Jens Ernstberger (@0xSerious) profile picture

agent infra @kontextsecurity

Joined April 2021
Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)Photo by Jens Ernstberger (@0xSerious)
It’s time to rethink authentication. ernstberger.xyz/posts/delega…
3
6
4
42
12,566
Redacting conclusions about risk whilst "focusing on risk assessment for the time being" seems quite contradictory, given the nature of an independent evaluation. Very curated information disclosure.
Replying to @METR_Evals
We agreed to pilot a compute allocation assessment, but Anthropic canceled this portion of the pilot (before we came to any substantial conclusions). Anthropic stated this was to keep the focus on risk assessment for the time being, and that it would revisit this later on.
1
207
🤨
Replying to @METR_Evals
1. Benchmarks provide one piece of evidence for acceleration. Since Oct 2025, AECI scores for Anthropic’s models have grown at 1.9x [95%: 1.5–2.8] the earlier rate. Analysis choices affect this estimate, and it remains unclear how benchmark scores map to real-world capabilities.
3
230
Just because I have an idea and write it down in a github repo (technically Claude did), it doesn't mean that the idea is original. Ideas are easy, execution is hard. Especially in a field with decades of already existing research.
anthropic just stole my best friends research he spent the last 10 months developing zero knowledge proofs for AI accountability, only for @AnthropicAI to announce remarkably similar research days after he applied for compute support. Can someone at Anthropic at least let him collaborate instead of jacking his research and calling it their own? proof attached below
1
10
686
If you send me an AI message, don't be surprised if you receive an AI message back.
1
5
303
cybersecurity awareness month ❤️‍🔥
6
191
Jens Ernstberger retweeted
Today, we’re killing the AI assistant Introducing mio.xyz, the first AI EMPLOYEE 𝘆𝗼𝘂𝗿 𝘄𝗵𝗼𝗹𝗲 𝘁𝗲𝗮𝗺 𝘀𝗵𝗮𝗿𝗲𝘀 In beta since July, Mio saved teams 1,000s of hours and completed 10,000s of tasks You don't need another agent. You need a 𝘀𝗵𝗮𝗿𝗲𝗱 one with your team Hire Mio today, $100 free credits for the first 100 companies
150
91
161
979
1,199,242
Why not 65%? I bet codex came up with that number, vibe extinction.
BREAKING: OpenAI recursive self-improvement researcher warns there is a 70% chance of “human extinction” in 3 years unless there is a coordinated slowdown between labs.
5
459
Agents can now solve complex math problems no human in our own company can solve, but sure let's wire up and open-weight model with a self-coded sandbox to keep token cost low, and for sure nothing bad will happen.
1
2
214
Turns out this aged well faster than I expected
The most important lesson from the OpenAI/Hugging Face incident isn’t that an agent escaped a sandbox. It's also not that the sandbox was "misconfigured". It’s that many “isolated” agents became one persistent actor. Each run was short-lived, but many shared a credential and could write to the same Artifactory instance. One agent left knowledge behind; another picked it up. For OpenAI, it was Artifactory. It could just as easily be Gmail, Linear, LinkedIn, logs, or a filesystem. You can isolate every process and broker every credential, yet still create a distributed actor through individually permitted reads and writes. If agents share state, security need share state too: detect across runs and kill access across the entire fleet.
1
3
377
The most important lesson from the OpenAI/Hugging Face incident isn’t that an agent escaped a sandbox. It's also not that the sandbox was "misconfigured". It’s that many “isolated” agents became one persistent actor. Each run was short-lived, but many shared a credential and could write to the same Artifactory instance. One agent left knowledge behind; another picked it up. For OpenAI, it was Artifactory. It could just as easily be Gmail, Linear, LinkedIn, logs, or a filesystem. You can isolate every process and broker every credential, yet still create a distributed actor through individually permitted reads and writes. If agents share state, security need share state too: detect across runs and kill access across the entire fleet.
3
1
6
835
seems like cloud agents, and not local cli, is the future. Thing is that they also need shared context, GOOD memory, connected to ALL your services (especially logs), etc - haven't found one that is quite there yet though
3
257
Time to speed up.
An open letter for a global surge in cyber defense, signed by over 100 organizations including Anthropic, AWS, Google, Microsoft, OpenAI, and Oracle.
3
420
Introducing 🦅 ShellRisk-Bench - a benchmark for evaluating the cyber risk of agent tool calls. As opposed to other benchmarks, ShellRisk-Bench is sourced from real agent trajectories and adversarial corpora, built to test guardrails at the base rates production traffic actually sees. Here's what we found ↓
13
8
2
34
1,948
On top of ShellRiskbench we built 🦅Kestrel - an open-source, on-device classifier for agent tool-call risk. – 0.947 precision, 0.922 recall on ShellRisk-Bench - beating the best frontier model by ~0.40 F1 – 21.5 µs mean latency per command (p99: 92 µs) - ~60,000x faster than frontier LLMs – Under 1MB, runs entirely on-device and commands never leave your machine
1
4
119
Someone told me AI presentation tools save them hours. And I saw the output which was no abstraction, tiny text, generic insights. The art is in simplification, not generation. Content is now a commodity. What's scarce is systematizing ideas in a way that is easily digestible, and drawing connections only you can see. A sub-par representation is worse than none at all.
2
15
418