PacSpace makes AI actions provable.
Joined August 2025
- Tweets111
- Following20
- Followers2
- Likes59
A question that used to be an investigation becomes a simple lookup.
A lab’s agents attacked @huggingface. It sifted 50 petabytes of logs and shared a selection with @METR_Evals and @redwood_ai. Some agents tried to tamper with them.
The record must exist.
PacSpace retweeted
Contain, detect, prove.
Planes have ID. Planes have real-time monitors. Planes aren’t allowed to touch the sky without a flight recorder.
This week, PHP flaws that had sat in the code for years were turned into working exploits with frontier AI.
The same day, a poisoned release of LiteLLM, the open-source gateway many teams route AI calls through, went out built to steal cloud and AI credentials.
It doesn’t replace containment or detection. It’s what you rely on when those fail.
Contain. Detect. Prove. The record must exist.
pacspace.io/
Credit to the researchers at @wiz_io and @calif_io , and to the @ThePHPF and @Cloudflare, for a coordinated fix.
Wiz’s write-up of PHP2Shell: wiz.io/blog/php2shell-php-rc…
Before an AI agent acts for your company, answer five questions.
Most teams can answer the first two. Very few can answer the last three, because they all need a record that exists when the action happens and shows any later change.
PacSpace makes AI actions provable.
ALT 5 point checklist before taking agent actions
Oct 2026: four labs breached, a government hacked, criminal charges, insurers writing exclusions.
Dec 2025: @DavidNgene wrote “they leave behind logs, not truth.” The evidence layer he described is what @PacSpace built.
We call it recordation. The record must exist.
pacspace.io/
Original article from December 2025: medium.com/@davidngene/agent…
Field Note updates for today's AI: pacspace.io/field-notes-ai-a…
PacSpace retweeted
@satyanadella: every significant model action should leave “tamper-proof human readable evidence.”
He’s right. But if the company running the agent holds the only record, that’s not evidence. It’s a claim.
The record has to sit where the other side can check it. That’s what we build at @PacSpace_io .
PacSpace retweeted
We’re beginning a process of publishing more frequent reports on model behavior, beyond what appears in our system cards and regular risk reports.
Today’s report describes four types of behaviors we’ve identified during evaluations and internal use. In each, Claude acted on real websites or systems in ways we didn’t intend, sometimes by working around a restriction instead of stopping.
All cases had minimal real-world impact. From an alignment and security perspective, we consider these behaviors significantly less severe than the cybersecurity incidents we reported in July and September.
Read the full report: anthropic.com/research/inves…
PacSpace retweeted
On Sept 29, six AI companies signed the White House accord and agreed to bring in outside auditors.
An outside audit is only as good as its evidence. Today the record of what an AI did lives with the company that ran it.
@PacSpace_io makes AI actions provable. Each action is recorded as it happens, nobody can change it, us included, and anyone can check it.
@elonmusk wants AI that is maximally truth-seeking. The record of what it did should be too.
Automated monitors face the same problem. A monitor can be fooled by what it’s watching, or start to side with it, and the lawsuit, the insurance claim, the liability finding that follows all rest on a log that a fooled monitor produced.
Here’s what the record shows. Agents found a way to cheat their tests within four hours. The real work came after: fooling the scoring program, altering the logs meant to catch them, manipulating the checks reviewing their work.