Tag: #ai

Researchers Detail 'TrapDoor' Supply‑Chain Attack That Hid Malicious Instructions in AI Assistant Files

Security researchers say 'TrapDoor' used malicious npm, PyPI and Crates.io packages to target crypto wallets and developer credentials, hiding commands in AI assistant files.

#cybersecurity, #supplychain, #cryptocurrency, #ai, #devtools

Meta says AI system RADAR has auto-reviewed 535,000+ code changes and landed 331,000

Meta details RADAR, an automated, risk-based code-review pipeline that reviewed 535,000+ diffs and landed 331,000+, per an arXiv preprint.

$META, #meta, #ai, #software, #code-review

Samsung Begins Shipping Samples of 12‑Layer HBM4E Memory to Major Customers

Samsung says it has started shipping samples of a 12‑layer, 48GB HBM4E stack to major customers, touting higher bandwidth and efficiency.

$005930.KS, #samsung, #hbm4e, #semiconductors, #ai

Preprint Finds LLM Watermarks Can Be Undermined by Supply-Chain PRNG Tampering

An arXiv preprint finds LLM watermarking can be bypassed by tampering with the inference PRNG; authors say quantum RNGs may defend.

#ai, #watermarking, #prng, #cybersecurity

Bank of Canada staff hold Canada neutral-rate estimate; U.S. range raised on productivity, AI

Bank of Canada staff keep Canada's nominal neutral rate at 2.25–3.25% while raising the U.S. estimate to 2.50–3.50% on stronger productivity and AI.

#canada, #neutralrate, #bankofcanada, #ai

New arXiv Preprint Says Adaptive 'Metis' Jailbreak Achieved High Success Against Multiple LLMs, but Results Aren't Yet Reproducible

An arXiv preprint describes Metis, an adaptive jailbreak that reportedly reached ~89% success across 10 LLMs, but its results are author-reported and not yet reproducible.

#ai, #security, #jailbreaking, #arxiv

Preprint Claims Perfect Scores by Multi‑Agent AI on Multiple Math Competitions; Results Unverified

A May 19 arXiv preprint says STAR‑PólyaMath scored perfect on AIME and Putnam and topped several benchmarks, but results remain unverified.

#ai, #arxiv, #math, #benchmarks

Google makes Search proactive with Gemini 3.5 Flash and new 'Search agents'

At I/O, Google made Gemini 3.5 Flash the default in Search AI Mode and unveiled “Search agents” that monitor the web and can complete tasks and bookings.

$GOOGL, $GOOG, #google, #search, #ai, #generativeai

Google makes Gemini 3.5 Flash the default across Search and Gemini app, unveils Spark agent

Google rolls out Gemini 3.5 Flash globally as the default in the Gemini app and Search AI Mode, unveils Spark agent and previews Gemini 3.5 Pro.

$GOOGL, $GOOG, #ai, #google, #gemini, #generativeai

Blackstone and Google form U.S. joint venture to sell Google TPUs as cloud service with $5 billion backing

Blackstone and Google will create a U.S. JV with an initial $5B equity commitment to offer Google’s TPU AI chips as cloud compute starting in 2027.

$BX, $GOOGL, #blackstone, #google, #ai, #tpu, #cloud

NSF Announces X-Labs, Pledges Up to $1.5 Billion Over a Decade to Translate Research into Technology

NSF unveiled X-Labs, a TIP-led program to fund milestone-driven teams with up to $1.5B over 10 years to build sensing and quantum platforms.

#nsf, #xlabs, #quantum, #ai, #innovation

Preprint Says AI Model SU-01 Reached ‘Gold-Medal-Level’ on Olympiad Problems — Results Not Independently Certified

An arXiv preprint says model SU-01 attains gold-medal-level on IMO/USAMO/IPhO problems, but reported scores lack independent certification.

#ai, #olympiad, #machinelearning, #reasoning

Google puts Gemini AI at the heart of Android, rolling out to Galaxy and Pixel phones this summer

Google will embed Gemini Intelligence into Android starting this summer on new Galaxy and Pixel phones, adding multi-step automation, smarter Chrome browsing and expanded Autofill.

$GOOGL, $GOOG, #google, #android, #ai, #gemini

Preprint Claims PRNG 'SeedHijack' Could Steer LLM Outputs; Authors Propose QRNG Defense

ArXiv preprint claims attackers can steer LLM outputs by hijacking sampling PRNG seeds; authors propose QRNG-based defense, but results are unreviewed.

#ai, #security, #prng, #qrng

Preprint Claims Nearly 147,000 ‘Hallucinated’ Citations in 2025 After Audit of 2.5M Papers

An arXiv preprint reports nearly 147,000 non-existent citations in 2025 after auditing 2.5M papers; findings come from an unreviewed abstract and are unverified.

#research, #ai, #arxiv, #scientificintegrity