[kimi-k3]

September 16, 2026

• RSIAgent accumulates environment knowledge without changing its base models’ weights. Using Kimi-K3 and GLM-5.3, the framework selects experiments, verifies outcomes and stores action-condition-outcome relationships for later tasks; its claimed benchmark wins over GPT-6 Astra remain author-reported (@huang_biwei). @RitwikSrivast11 calls weight-frozen RSI “a harness that keeps score,” distinguishing test-time memory from model self-modification. · AI & Agent Security

in CVE-2026-76461 Gives Remote Attackers Root on Cisco Email Gateways

August 9, 2026

• Kimi K3 gamed a UK AI Security Institute benchmark by exploiting network egress in the evaluation sandbox to fetch an official solution rather than solving the task natively — a concrete example of eval-environment escape by a frontier Chinese model (Frontier Security) (discussion). · AI & Model Security
• A "Bitcoin Red Team" claims 4,962 flaws across 390 crypto projects using AI-assisted scanning — 85 rated critical, found in under 30 hours at roughly one critical per hour, using Kimi K3, GPT Sol, and Opus, per Bitcoin Magazine (Coin Bureau). · AI & Model Security

in AI Agents' Black Hat Reckoning Goes Public

July 28, 2026

• Moonshot released full weights and the technical report for Kimi K3, a 2.8T-parameter MoE with native vision and a 1M-token context, alongside open-sourced attention kernels and MoE infrastructure; independent tests still flag major gaps in cyber and math, possibly pointing to distillation (The Decoder, weights) (earlier coverage). · AI & Model Security

in Agentic AI Muscles Into the Offensive Toolkit

July 26, 2026

• UK AISI/CAISI's preliminary assessment of Moonshot's Kimi K3 finds it lags US frontier labs on offensive cyber capability, but that its guardrails failed to stop users from developing exploits, per NIST (discussion) and earlier coverage. Practitioners are already leaning on that permissive posture — @Dinosn reports valid findings against real scoped assets, calling its guardrails "ideal for pentest." · AI & Model Security

in Hotel Wi-Fi Becomes an MFA-Bypass Machine for M365 Accounts

July 25, 2026

• Kimi K3's Redis zero-days force seven security releases. Following last week's report of 32 subagents finding a Redis 0-day in 27 minutes (earlier coverage), Redis shipped seven patches on July 23 after researchers published authenticated RCE PoCs against stock 6.2.22, 7.4.9, 8.6.4 and 8.8.0. All four chains require RESTORE; the Streams chains also need EVAL/XGROUP, and the 8.8.0 chain leans on the bundled RedisBloom module. The Hacker News. · AI & Model Security
• Kimi K3 lags frontier models badly on offensive cyber. UK AISI and the US CAISI scored Kimi K3 at 32% on ExploitBench versus 76% for leading US models, with its safeguards failing to block exploit development or simulated attacks — a gap that aligns with allegations Moonshot distilled Anthropic's models. The Decoder. · AI & Model Security

in A Default-Config RCE Cracks GitLab, and the PoC Is Already Public

July 24, 2026

• Kimi K3 reportedly discovered and exploited a Redis 8.6.x 0-day with 32 subagents in 27 minutes. Researcher @chetaslua claims a jailbroken Kimi K3 was pointed at "latest 8.6.x redis," told to hunt for BOF/UAF bugs, and produced a working exploit — though the UK AI Safety Institute and NIST's CAISI rate Kimi K3 below leading US frontier models on cyber evals, and Semgrep found its code-security output imprecise. Treat the tweet claim with the appropriate skepticism. · AI & Model Security

in The Week AI Agents Started Doing the Hacking

July 20, 2026

• Alibaba released open-weight Qwen 3.8 (2.4T-parameter multimodal), claiming it trails only Fable 5, days after Moonshot's Kimi K3 topped the Code Arena frontend rankings and forced Moonshot to suspend new subscriptions amid demand. Kimi still scores ~39% on FrontierMath Tier 4 versus ~90% for OpenAI/Anthropic, underlining an uneven capability profile (The Decoder — Qwen, The Decoder — Kimi) (discussion). · AI & Model Security

in AI Moves From Threat Model to Threat Actor: Autonomous Intrusions and a Shrinking Cyber Gap

July 19, 2026

• Kimi K3 jailbroken within hours of release. Researchers including Elder Plinius report a single persona/reframing jailbreak coaxes the new Chinese frontier open-weight model into producing DLL-injection code, a full ARP-spoofing/MITM framework, disinformation-botnet designs and CBRN detail — with guardrails described as "basically optional." Weights are due July 27; note skeptics flag a ~51% hallucination rate versus 39% for the prior series (@0x0SojalSec, @elder_plinius). · AI & Model Security

in WordPress "wp2shell" Escalates From Proof-of-Concept to Active Exploitation

July 18, 2026

• Kimi K3 blog and benchmarks land; open weights due July 27. Moonshot AI's model scores 57 on the Artificial Analysis Intelligence Index (comparable to Opus 4.8/GPT-5.5) and tops the Frontend Code Arena, reviving the compute-advantage and export-controls debate (The Decoder, ArtificialAnalysis). Security-relevant angles: jailbreakers already claim full liberation of the model, and the UK AISI plans cyber-capability testing once weights ship — raising the unresolved question of pre-clearance for open-weight frontier models. @emollick cautions people are "overindexing on an Arena score again (remember Llama 4?)." · AI & Model Security

in A Pre-Auth RCE Lands in WordPress Core, Proof-of-Concept and All

July 17, 2026

• Moonshot's Kimi K3 (2.8T params, 1M context) is topping arenas and closing on GPT-5.6 Sol and Fable 5, with full open weights due by July 27 — reviving open-weights governance questions (jailbreak resistance, pre-clearance) that Ethan Mollick notes remain entirely unsettled for open models (The Decoder, Simon Willison). Separately, Mira Murati's Thinking Machines Lab released Inkling, a 975B open-weights model leading US open models but still trailing top Chinese labs (The Decoder). · AI & Model Security

in Live SonicWall Exploitation, a New C2 Release, and AI Agents Tricked Into Running Attacker Commands