Dhruv AHUJA
@new23d.bsky.social
Making network egress filtering effective, reliable and usable. Founder & Chief Engineer at @chasersystems.bsky.social Blog: https://www.new23d.com/
says No Internet on the bottom
Chaser Systems@chasersystems.bsky.social · 5d agoCome grab a pair! We've gone all in on socks this year for @fwdcloudsec.org EU 2026 🧦 Various sizes of ethically made, high quality socks, produced by the excellent people at @Stand4Socks Every sock has a sister pair that has been donated to a rough sleeper.
I hope there'll be stickers left for me and I won't have to use one of the 4 earlier iterations of them.
Chaser Systems@chasersystems.bsky.social · last wk.Come grab some free socks & stickers from us at @fwdcloudsec.org EU London 2026. The team are hard at work preparing the stall, fiddling with electronics 👨🔬 and trying the socks on for size 🧦
My latest blog on the Agent's DNS tricks in the HF - OAI incident got picked up by @marcolancini.it's newsletter issue 352 😃 Newsletter link: cloudseclist.com/issues/issue...
Excellent guidance by #NCSC on controls for LLM-driven Agents. #DiscrimiNAT #OTF by @chasersystems.bsky.social can help you achieve Level 2 and 3 of 'Restrict network access'. Get in touch for a demo or search #DiscrimiNAT #OTF and book one. www.ncsc.gov.uk/blogs/managi...
Managing the cyber risk of agentic AI
Use safeguards, sandboxing and active oversight to realise the benefits of autonomous systems while limiting the unintended activity.
ncsc.gov.uk
The 1968 2001: A Space Odyssey is still a benchmark in film-making.
I wrote a blog post on how and why the Agent manipulated DNS in the Hugging Face – OpenAI incident.
Chaser Systems@chasersystems.bsky.social · 3w agoThree variations on subversive use of DNS by the Agent are documented in Hugging Face's technical writeup of the July 2026 security incident involving OpenAI models. In this article, @new23d.bsky.social discusses what each of these three types of DNS workarounds... chasersystems.com/blog/the-cur...
'Security is architecture' comes to mind. If Artifactory's access to the internet was also filtered, the Hugging Face - OpenAI incident could've been prevented. At @chasersystems.bsky.social this is why we built DiscrimiNAT. It even detects DNS tricks by Agents like resolution overrides.
I first thought of getting a ham radio licence during '20-21😷. Never got around to it even though I bought the study guides then. Passed the Foundation Level last week and today was issued a callsign, M7XIT, by Ofcom to transmit at 25W. Don't have a transceiver so don't CQ yet🫠
eval'ed gpt-5.6-luna & deepseek-v4-flash-0731 vs my prev fav mimo-v2.5-pro today. gpt preferred using python for tool calling & subtasks; the others bash/gnu cli. gpt also had a greater breadth of understanding the natural language. gpt-5.6-luna is my new fav for a few days!
Does anybody know an illustrator/graphics designer who can tweak and convert AI-generated two sticker ideas I have to print-ready? Theme is computer circuits, LCD screens and a bit retro Windows 95 dialog boxes. Need this done asap, really. Preferably Cambridge/London based.
This isn't a reason to be complacent. We're working on an Agentic Defence Assurance and Product Integrity Testing (ADAPT) harness to continually challenge our assumptions and discover new evasions. More to be released in a few weeks on that. Outbound Traffic Inspection (OTF)...
Chaser Systems@chasersystems.bsky.social · last mo.Happy to reassure our users that the #egress evasion TTPs used by the Agent in the OpenAI & HuggingFace incident have long been mitigated in #DiscrimiNAT OTF (outbound traffic filtering). In fact in our demo, we show these attacks being caught. #2 is SNI spoofing, btw huggingface.co/blog/agent-i...
"security systems architecture experts" TIL about MITRE D3FEND: Detection, Denial, and Disruption Framework Empowering Network Defense. It's updated quarterly and useful to map "defensive cybersecurity techniques" to TTPs. Nice one. d3fend.mitre.org
MITRE D3FEND Knowledge Graph
D3FEND is a knowledge base of cybersecurity countermeasure techniques. In the simplest sense, it is a catalog of defensive cybersecurity techniques and their relationships to offensive/adversary techn...
d3fend.mitre.org
Update profile on APT28/Unit 26165/Fancy Bear by on Gov·UK [1]. "Unit 26165 accessed private IP cameras near military facilities, ports, train stations and border crossings in Ukraine, Moldova and 11 NATO countries to track the movement of foreign assistance" They seem to 1/3
65% of breaches could've been prevented according to this research by Niels Provos if only Hardware second factors, Egress control and Positive execution control would've been in place. I couldn't agree more when he says, 1/2 securityblueprints.io/posts/three-...
Three Security Invariants Could Prevent 65% of Breaches: Analyzing 70 Incidents and Building CISO Challenge
Analysis of 70 real-world data breaches reveals that three security invariants — hardware second factors, positive execution control, and egress restrictions — could have prevented 65% of incidents.
securityblueprints.io
VPC Endpoint Bucket Policies are completely ineffectual in preventing exfiltration if the workload has access to the Internet. Just upload to the $otherCloud's storage endpoint. This is where #egress filtering at the NAT gateway layer prevents a compromise.
Interestingly, in Salesforce, IP ranges security is only applied at login time at Org or User Profile/Session levels by default. Once the token is issued post login, it can be used outside of the IP ranges perimeter. Unless one ticks ✅ the additional 1/2
Well, OAuth didn't solve the problem did it: "The attackers pushed a code update capable of collecting OAuth tokens Klue’s customers use to connect Klue to their own systems." [1] 1/4 [1] www.huntress.com/blog/klue-br...
There is a computational, time and environmental cost to #PQC in TLS... Google enabled it server-side on Feb 12 and any OpenSSL v3.5+ or Go crypto/tls v1.24+ client since then is sending ~1200 bytes of additional session key data than before. This takes the TCP packet of the 1/4
Using VPN to connect to an already TLS endpoint doubly encrypts the data in transmission. Sheer waste of resources and creating a bandwidth bottleneck for all users, IMO. If it's a "private" network range that is being routed to via the VPN, there's a case, yes. But think of 1/3
Supremely excited about we're building next 🚀 This is a foundational first step towards the infra needed 🏗️
Chaser Systems@chasersystems.bsky.social · 4mo agoWe were just assigned a /29 #IPv6 block by RIPE. We now have 633,825,300,114,114,700,748,351,602,688 IPv6 addresses. inet6num: 2a05:6340::/29 netname: UK-CHASERSYSTEMS-20260518 country: GB org: ORG-CSL88-RIPE admin-c: AA44781-RIPE apps.db.ripe.net/db-web-ui/lo...
Our #egress filtering solution for containing agent behaviour in the cloud appears to be hardened enough as of today 🤞 . Of course, it is not a system you log into or get shell access during the course of work. And all input is first sanitised by the cloud APIs and then our own logic.
Chaser Systems@chasersystems.bsky.social · 4mo agoDiscrimiNAT v2.40 shipped yesterday on GCP and day before on AWS. It has Canonical-supplied mitigations applied for Copy.Fail. It is also hardened to the CIS standard and therefore the exploits for Dirty Frag fail to work on it. We are monitoring the situation and will push an update if needed.
"Weekends are the untapped frontier. Still room to scale." [1] Superb insight on the Show HN thread for a satirical GitHub outages contributions chart 🟥🟧🟨 😇 [1] news.ycombinator.com/item?id=4803... [2] red-squares.cian.lol
Agentic red teaming (or malicious activity) have lowered the bar for their human actors when it comes to #evasion around #egress #filtering tech; and raised the bar for vendors making such tech 🙋♂️ <insert bittersweet meme>
GitHub down again. I guess I will finally have dinner at a sensible time today 😅 www.githubstatus.com/incidents/ql...
