Georgia Tech

Headlines · 5

1. 3d
   Agent MechSuits Adds Internal Safety Controls to CLI Agents

   Researchers from Georgia Tech, Rutgers, the University of Southern California and other institutions proposed Agent MechSuits, a NeurIPS 2026-accepted method that reads hidden-state safety signals before each command-line agent action and intervenes on 10 dimensions without updating the main model. Across LLaMA-3.1-8B, Qwen-2.5-7B and Gemma-2-9B, the reported MAS current-step detection accuracy averaged 97.32%, while Qwen-2.5-7B's malicious-step ASR fell from 91.30% to 4.35%. The method also had limitations: generation-collapse rate rose from 2.24% to 18.54%, the evaluation used language-model judging, and the probe did not reliably detect prompt injection arriving through environment feedback.

   新智元 · 🔥 0
2. 5d
   Nagare Labs advances Z2RAM memory-on-compute prototype

   Nagare Labs says its Z2RAM could place SRAM-speed memory with DRAM-like density directly on AI accelerators, addressing bandwidth and heat constraints that make stacked HBM difficult. Founder papers report a three-transistor cell retaining data for more than 300 seconds at 85°C, while Gen 1 targets such as 5× SRAM density and 14× DRAM speed remain vendor projections. The technology remains a lab prototype without a fabrication facility, and no published cell has demonstrated retention above 95°C.

   Tom’s Hardware · 🔥 0
3. 6d
   AI Is Finding Patterns in Dolphin Communication

   WIRED AI reports that the Wild Dolphin Project is using AI to identify patterns in dolphin vocalizations, working with researchers at Georgia Tech and Google DeepMind. The team has also used an underwater wearable computer to play dolphin sounds associated with objects and habitats in an attempt at two-way communication. Researchers caution that mimicry does not establish understanding, and full translation remains out of reach.

   WIRED AI · 🔥 0
4. Oct 31
   Anthropic warns small AI models could enable self-replicating botnets

   Engadget reports that Anthropic's Frontier Red Team found small AI models may be only months away from replicating and spreading across insecure devices, creating slow, hard-to-kill autonomous botnets. Experts broadly agreed that such capability could emerge in the near future, but Georgia Tech professor Michael Bailey disputed the characterization of the emulation as full self-replication. He said it demonstrated sharding and device hopping in a network where the exploit was trivial, rather than worm-like propagation, and that Anthropic did not attempt model retraining. Anthropic suggested that internet service providers notify customers of infections and limit infected devices to a safe list of websites, as well as possible regulation. Bailey separately emphasized patching, device isolation, strong passwords and security hygiene.

   Engadget AI · 🔥 0
5. Oct 27
   ORISA recovers deleted local-AI chats from Android memory

   Georgia Tech researchers developed ORISA, a forensic tool that reconstructs deleted conversations from AI apps running locally on Android by searching device memory for residual tokens and KV caches. In an experiment using a note-keeping assistant built with ChatterUI, the tool restored a working session containing 26 of 100 fictional facts, while the original session retained access to 14. The technique requires forensic access to device data, and recovery becomes less reliable as newer activity overwrites memory.

   ZDNET AI · 🔥 0

Experience and discussion from the community

Share my Georgia Tech experienceAsk about Georgia Tech