AISecHub

Kanalga Telegram’da o‘tish

Ko'proq ko'rsatish

Mamlakat belgilanmaganTexnologiyalar & Aralashmalar28 160

2 478

Obunachilar

+224 soatlar

+197 kunlar

+14630 kunlar

319

Post ko'rishlar

~ 16524 soatlar

~ 19848 soatlar

12.90%

Muloqot nisbati

~ 12

Kuniga postlar

Ads index

beta

Ma'lumot yuklanmoqda...

O'xshash kanallar

Ma'lumot yo'q

Muammo bormi? Iltimos, sahifani yangilang yoki bizning qo'llab-quvvatlash boshqaruvchimizga murojaat qiling>.

Taglar buluti

Kirish va chiqish esdaliklari

---

Obunachilarni jalb qilish

Iyun '26

+86

3 kanalda

May '26

+163

4 kanalda

Get PRO

Aprel '26

+100

2 kanalda

Get PRO

Mart '26

+192

3 kanalda

Get PRO

Fevral '26

+239

5 kanalda

Get PRO

Yanvar '26

+394

4 kanalda

Get PRO

Dekabr '25

+204

4 kanalda

Get PRO

Noyabr '25

+140

4 kanalda

Get PRO

Oktabr '25

+167

3 kanalda

Get PRO

Sentabr '25

+270

1 kanalda

Get PRO

Avgust '25

+95

1 kanalda

Get PRO

Iyul '25

+93

2 kanalda

Get PRO

Iyun '25

+68

1 kanalda

Get PRO

May '25

+375

3 kanalda

Get PRO

Aprel '25

+43

2 kanalda

Get PRO

Mart '25

+735

1 kanalda

Sana	Obunachilarni jalb qilish	Esdaliklar	Kanallar
21 Iyun	+7
20 Iyun	+2
19 Iyun	+2
18 Iyun	+7
17 Iyun	+6
16 Iyun	+12
15 Iyun	+2
14 Iyun	+2
13 Iyun	+6
12 Iyun	+1
11 Iyun	+1
10 Iyun	+7
09 Iyun	+4
08 Iyun	+4
07 Iyun	+1
06 Iyun	0
05 Iyun	+2
04 Iyun	+5
03 Iyun	+6
02 Iyun	+4
01 Iyun	+5

Kanal postlari

https://signals.forwardfuture.ai/loop-library/

2	https://medium.com/@taleliyahu/ai-red-teaming-will-need-qualified-human-pentester-6a0beecd2aa0 The tool will not be enough. AI red teaming is becoming one of the default rituals of enterprise AI security. That sounds like progress. It is progress. But the industry is already making the mistake we always make when a new control category appears: we turn a hard assurance problem into a tool problem. That is not where this ends.	202
3	CVE-2026-55837: dbt MCP Server OAuth context endpoint leaks dbt Platform tokens The advisory describes a local OAuth helper FastAPI server in dbt-mcp exposing GET /dbtplatformcontext without authentication or host/origin validation, allowing retrieval of dbt Platform tokens via that endpoint. #MCP #AgentSecurity #AISecurity #Advisory https://github.com/advisories/GHSA-jr33-mw75-7j8f	230
4	AutoJack: Web Page Hijacks an AI Browsing Agent to Reach Host Code Execution Reported exploit chain: a single attacker-controlled page can use JavaScript to cross the browser/agent trust boundary by talking to a privileged local service and spawning a host process, after the user merely steers the agent to that page. #AgentSecurity #LLMSecurity #AISecurity #News https://thehackernews.com/2026/06/autojack-attack-lets-one-web-page.html	223
5	A Red-Teaming framework for safety evaluation of LLMs specialized in the financial sector https://arxiv.org/abs/2606.19887v1	214
6	gdm-ai-control-roadmap.pdf	273
7	Securing the future of AI agents - https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/securing-the-future-of-ai-agents/gdm-ai-control-roadmap.pdf \| https://deepmind.google/blog/securing-the-future-of-ai-agents/	265
8	CVE-2026-0755: gemini-mcp-tool OS command injection and @file exfiltration via prompt quoting The advisory describes a trust-boundary failure where untrusted prompt text is interpreted by the Gemini CLI as @file includes (reading local files), and on Windows can be parsed by cmd.exe when quoting is missing, enabling metacharacter-based command injection. #MCP #AgentSecurity #AISecurity #Advisory https://github.com/advisories/GHSA-4h5r-5jm8-jxjm	279
9	Defensive Misdirection Against Model-Guided Automated Attacks on Agentic AI Systems The paper models an attacker using automated judges and search to refine jailbreak prompts, arguing that predictable detect-and-block refusals can become a feedback signal as query budget grows. It evaluates a detect-and-misdirect strategy that returns safe but intentionally misleading outputs to degrade the attacker judge and bound asymptotic success rates. #AgentSecurity #LLMSecurity #AISecurity #Research https://arxiv.org/abs/2606.20470	243
10	NRT-Bench: multi-turn red-teaming benchmark for LLM agents in a safety-critical control-room simulation NRT-Bench evaluates LLM “operator” agents under adaptive, multi-turn message injection where failure is measured by objective loss of a critical safety function, not LLM-judged text. Results suggest robustness is model- and defense-stack-dependent, with largely non-overlapping failure cases across tested models. #PromptInjection #LLMSecurity #AISecurity #Research https://arxiv.org/abs/2606.20408	220
11	Layered Security Framework Against Prompt Injection in RAG Chatbots This paper proposes a 3-layer RAG defense that gates user input, enforces an instruction/provenance hierarchy during context assembly so retrieved text can’t override policy, and audits outputs with rules plus semantic-drift checks, with a feedback loop from structured logs. #PromptInjection #LLMSecurity #AISecurity #Research https://arxiv.org/abs/2606.19660	212
12	Matn yo'q...	227
13	Matn yo'q...	244
14	Matn yo'q...	46
15	Already posted — ▉	1
16	Matn yo'q...	46
17	All 9 items posted to AISecHub. One ▉	1
18	Matn yo'q...	45
19	AI-Accelerated Vulnerability Discovery and Patch Debt The piece frames a scaling mismatch: AI-assisted vulnerability discovery can increase finding throughput faster than organizations can validate, prioritize, and deploy fixes, expanding "patch debt" as an operational bottleneck rather than a purely technical one. #VulnerabilityResearch #AppSec #AISecurity #Report https://labs.cloudsecurityalliance.org/research/ai-accelerated-vuln-discovery-systemic-patch-debt-v1-0-csa-s	257
20	Last one timed out — retrying. ▉	1

Barcha postlarni ko‘rish