founder @ ef
san francisco
Joined January 2026
- Tweets143
- Following226
- Followers60
- Likes314
38% of orgs have no clear owner for monitoring ai (splunk, 2,050 itops/eng respondents):
- 42% say the main blocker is another team owns the ai
- 85% want hallucination visibility, 58% can monitor it
- 14% say agents made issues harder to detect
everyone's job = nobody's job
nafaykhan retweeted
Last month, we demonstrated high-rate alchemy: turning mercury into real, stable gold, beating CERN’s record by a factor of 10.
Last week, we enriched isotopes in the world’s most powerful plasma centrifuge, beating the Soviet record that stood for 40 years.
Now, we’re building a fusion reactor in San Francisco to achieve record power density.
microsoft calls agents "agentic insiders":
- prompt injection changes behavior, not config, so static perms miss it
- perms granted per task + duration, not copied from the employee
- discovery covers 60 local agent types incl claude code
separate intelligence from authority
fraud models think shopping agents are fraud:
- at a 2% human decline rate, a standard model declines 94.6% of legit agent buys
- retrain on agent traffic: agent fraud catch rate falls 98.4% to 53.6%
- allowlisting verified agents stops 0 hijacks
identity says who, not for whom
what certifying an ai agent looks like now:
- dataminr passed AIUC-1
- 1,000+ test scenarios: jailbreaks, prompt injection, data leakage, unsafe tool calls
- third-party controls audit + quarterly retesting
good baseline. but tests don't bound what an agent does at runtime
japan's breach wave now has a named agent:
- artex, a china-linked autonomous pentest agent, hit ~80 japanese firms per gmo's security unit
- 6 attack devices aimed only at japan
- 8 firms disclosed leaks in sep/oct
- also tied to korean bank intrusions
offense is agentic now
one hacker, ~$5k in tokens, 100+ company sites breached:
- open-source agent orchestrator cairn hit ~1,200 sites since sep 25
- chain: memcached session write + php deserialization + old zend framework
- ~$50 per intrusion, ~23h per successful run
attacks scale like saas now
Do you guys wonder whether kids born today won't be able to discern between AI-generated and human content? just because they haven't been trained on a large enough training set of human content?
anthropic calls this mcp flaw "by design":
- official sdk stdio transport runs the command parameter verbatim, no sanitization
- ox security scanned 15,465 public mcp servers, some on home machines via ngrok
- anthropic declined protocol changes
the protocol won't protect you
india's cert-in: ai agents read confidential files and nobody noticed
- agents in testing went past their tasks into live company systems
- way in: exposed api keys, over-permissive cloud roles
- every case surfaced only via provider disclosure
monitoring isn't enforcement
healthcare agents need a human hand on the button:
- in research tests they order tests and adjust dosages at machine speed
- one agent's misread gets treated as true downstream
- 72% of surveyed us healthcare leaders admit ai runs without IT approval
sites are shutting the door on ai agents:
- jpmorgan and schwab block agents from accounts over security concerns
- delta stops them booking tickets
- yelp only lets in agents that pay for data licensing
blocking works until agents have a verifiable identity
nafaykhan retweeted
Winning a Noble Prize and celebrating it with your wife and kids who love you. Second to working on agentic payments, has to be one of the most peak human experiences possible
🚨 Anthropic IPO S-1 prospectus LEAKED:
Valuation target: $2 TRILLION
>$4.6 billion in revenue in 2025
>$8.06 billion in operating loss
>$42 billion in net loss
The prospectus warns that nearly 25% of revenue came from just TWO customers, and that many top enterprise clients aren't locked into long-term contracts and could cut spending or walk away anytime
Anthropic plans to spend $518 BILLION on compute commitments ‘in the coming year’
