Opinions are my own. Professional Infosec Researcher and Adversarial AI thinker.
Joined March 2023
- Tweets104
- Following55
- Followers1.6K
- Likes77
REA is scary from a security perspective. Anyone can effectively RE Cobalt Strike, EDR tools/agents, Microsoft Defender. Got more niche interests? How about RE Chrome & Edge for faster 0day’s?
collusion.wiki - I spent all night doing a deep technical analysis. In short, the essentially utilized Domain Fronting to bypass sandbox controls. Larger datasets were staged on jsonhero.io (16 distinct stores) addressed by the target's exact schema (?path=regCF_county_2021…), with raw-export readback. Three of these stores were inspected directly — they are public — and each held authentic retrieved data (§8.5). The staging is two-layered, mirroring how the SEC map is assembled: the values (regCF_county_*) pulled from sec.gov through CORS proxies, and the geometry (us-ma-all.geo.json, the county shapes) pulled from Highcharts/GitHub — both parked on jsonhero and joined the way the report's own main.js joins them, on ["hc-key","code"]. Some of those datasets are still accessible today.
Erhm... After reading this, I'm pretty sure this is just the tip of the ice berg now. In the original report, the agents demonstrated implementing encryption for message integrity - if they (the agents) suspected much worse, they would most likely revert to covert comms. That's just opsec at that point.
A Cyber Guy retweeted
Not sure if this is a large scale AI outage, or if it's just a critical internet backbone provider issue at the moment like CloudFlare / AWS, but the top four on DownDetector are all top AI platforms.
I did get a cf-ray: number with CMH on the back of it in an AI product just now - it might be CloudFlare...
I wouldn’t expect a single fixed token signature. I’d expect a keyed, context-dependent statistical signal distributed redundantly across generation decisions—potentially conditioned on semantic context so edits degrade confidence gradually rather than destroying the mark outright.
For anyone wondering, with a little creativity:
Zero-width character sequences
Unicode variation selectors
Unicode tag characters
Invisible-character clustering
Soft-hyphen (U+00AD) placement
Periodic soft-hyphen patterns
Word-boundary soft-hyphen patterns
Keyword-breaking soft hyphens
Mixed-script homoglyphs
Unicode confusables
Latin/Cyrillic script mixing
Hidden HTML/CSS text
Raw-text vs. rendered-text differences
Unicode normalization differences
Encoded zero-width sequences
Printable-data decoding from invisible characters
Character placement/distribution patterns
Non-breaking-space anomalies
Trailing/structural whitespace patter
Just wait until it’s individual fingerprinting arxiv.org/abs/2602.16800
My harness I built inventories invisible Unicode and flags clustered runs of 4+ tag characters / variation selectors — characters that can exist in the underlying text while being completely invisible on screen.
I also designed a second-stage check that attempts to decode zero-width sequences. If the invisible characters resolve into printable data, that’s a very different signal from “this writing sounds like AI.
If you do it right. You can get Googles AI mode to actually run an idea in a sandbox for you. Pretty neat…I wonder what else it could do…
This bill does function more like an omnibus than a standard Continuing Resolution (CR). While CRs typically extend existing funding levels temporarily, this bill includes substantial new appropriations, policy changes, and provisions beyond simple funding extensions.
