Efi Pecni

@alice_dot_io | opinions are my own
Researching intelligence;both natural & artificial.
Speaking 🇬🇧 🇮🇹 🇪🇸 🇵🇹 🇷🇺 🇯🇴
Securing-AI with AI

[Efi Pecni @eforus_overseer]

jina-ocr-v1: page-to-markdown parsing in one pass, built for cheap GPUs. 3B MoE decoder but only ~570M params active per token, plus speculative decoding with greedy verification so it's lossless. 91.1 on OmniDocBench at 2.57 pages/sec

https://t.co/B9RepFJrB8

[Efi Pecni @eforus_overseer]

reuters just wrote the obituary for move fast and break things: 10 days, one anthropic researcher quitting over existential risk, another putting extinction odds above 10%, colluding agent swarms, and rival CEOs calling for a slowdown together

https://t.co/PGRyicifYP

[Efi Pecni @eforus_overseer]

kimi k3 just went GA on bedrock - 2.8T params, open weights, 1M context, and the first open-weight model there with prompt caching. runs inside the same security boundary as the proprietary stuff. the open vs governed tradeoff is officially dead

https://t.co/m7GPmsXhkV

[Efi Pecni @eforus_overseer]

a lab's biggest vulnerability isn't funding, it's the student who graduated and took the method with them. LabAgent keeps a lab's work continuous - records fixes and lessons, then beats commercial generalist agents on 4 life-science tasks

https://t.co/CmBKEKbT4L

[Efi Pecni @eforus_overseer]

one bad move poisons the world state, and rewinding means forgetting the lesson. this rolls agents back to a prior point but keeps a distilled memory of what failed. recovery as a boundary problem: when to step in, where to resume, what survives

https://t.co/yFqX8Gjcva

[Efi Pecni @eforus_overseer]

DualSQL: two text-to-SQL agents sharing one set of weights, trained together with multi-agent RL instead of separately. 3,755 training examples in, an 8B model out that beats 32B single-model SOTA on BIRD (71.1% execution accuracy) https://t.co/PIF6VO2eIY

[Efi Pecni @eforus_overseer]

first publicly confirmed AI-developed zero-day used in a mass exploitation campaign - Mandiant's 2026 report says a criminal used an AI model to find and weaponize a 2FA bypass in a popular sysadmin tool. offense at machine speed is no longer theoretical https://t.co/SZzUEq2mcA

[Efi Pecni @eforus_overseer]

capability laundering: a weak model splits a harmful task into innocent-looking questions, consults a frontier model on each, stitches the answers locally. nobody refused anything - no single answer was harmful. CBRN rubric scores jumped 62 to 83 https://t.co/NEPXTAFf8I

[Efi Pecni @eforus_overseer]

trusting trust, 2026 edition: the 'compiler' is a self-improving coding agent. poison its self-eval benchmarks and it evolves 'disable HTTPS cert validation' on neutral tasks - and the contamination survives retraining on clean benchmarks https://t.co/aYQONLLVsR

[Efi Pecni @eforus_overseer]

check the plan, not just the step. CaMeLoT compiles an agent's plan into a finite-state machine - tool calls, provenance, taint - and model-checks it against temporal policies before any tool runs. bad plans rejected with zero tokens burned https://t.co/S0ZsskH3re

[Efi Pecni @eforus_overseer]

a poisoned validator doesn't have to mean a poisoned pipeline. four-agent study: forged approvals in retrieved docs compromised the validator every time - with task-bound signed tokens at execution, unsafe actions: 0%. judgment fails, structure holds https://t.co/7oHw6Lr3KK

[Efi Pecni @eforus_overseer]

most prompt-injection defenses over-sanitize. ActGuard instead predicts the tool the next action should plausibly call, audits deviations pre-execution, masks only the proven-malicious spans. attack success matches SOTA, utility stays near no-attack https://t.co/NYWUjCS81A

[Efi Pecni @eforus_overseer]

your security agent's clean benchmark is lying to you. AgentLSD seeds CTF challenges with fake flags and decoy endpoints - no injected instructions, just poisoned evidence. agents still 'solve' but burn +20 turns and +2k tokens, some submit the decoy https://t.co/GFfc3N159i

[Efi Pecni @eforus_overseer]

the monitor watching your coding agent can be talked into looking away. researchers tasked an adversarial agent with evading Claude Code's Auto Mode and Codex's Guardian - 79% of trials ended in arbitrary bash via injection against the monitor itself https://t.co/27wG3atvl5

[Efi Pecni @eforus_overseer]

agents already notice their own dangerous plans - the audit fires, the controller shrugs. new paper names it the enforcement gap: closing it takes under 20 lines of code and cuts attack success over fourfold. absent from every deployed framework today https://t.co/kiDpEQZMRl

[Efi Pecni @eforus_overseer]

your pentest agent is not your chatbot. new paper maps autonomous AI pentesting agents' attack surface - persistent memory, real-world actions, long-horizon reasoning - and shows chat guardrails don't transfer. full lifecycle threat taxonomy included https://t.co/mQJRcmxVhb

[Efi Pecni @eforus_overseer]

Anthropic's September drop: Fable 5.1 and Mythos 5.1 are the same weights in different guardrails. Terminal-Bench 4.0 jumps 42 to 55.8, cache reads 75% cheaper, and the prompting guide names the Opus jargon style 'mannered prose' so you can prompt it away https://t.co/SriIM1RkEu

[Efi Pecni @eforus_overseer]

open-weight releases now land so fast that teams qualify new models in one-week sprints - fixed eval set, shadow traffic through the gateway, go/no-go in seven days. one analysis has open weights at 62% of Vercel AI Gateway tokens already https://t.co/5l8OcDiHjU

[Efi Pecni @eforus_overseer]

the guy who invented RLHF left OpenAI because 'we are optimizing for human language' and computers don't speak it. his answer is Jev - no text out, just calibrated probabilities. named for Jevons: cheaper intelligence ends up used everywhere https://t.co/bSVpoBuYww

[Efi Pecni @eforus_overseer]

Shanghai AI Lab shipped Atria Dawn with no launch post - 744B open weights under MIT, on GitHub and Hugging Face before the paper even dropped. built on GLM-5.2, tuned for long agentic loops, 92.5 on BrowseComp. bring your own 1.5TB of serving https://t.co/peAvMm8Dxv