All news with #llm security tag

🔒 A new research paper shows a flaw in how OpenAI, Anthropic, and Google carry encrypted reasoning between API calls, enabling recovery of hidden internal reasoning and secrets from session logs. The team demonstrated replay and decoding attacks that recovered API keys, passwords, and other private artifacts from publicly available agent traces. Vendors implemented mitigations and the authors say the main extraction no longer reproduces as of August 2026, but developers are urged to strip opaque reasoning blocks from shared logs.

🛡️ Researchers from Tracebit report that embedding prompt injections alongside secrets stored on AWS can disrupt AI hacking agents by triggering LLM guardrails. These injected prompts instruct the model to perform forbidden actions, causing the LLM to shut down or stop following prior commands—a technique the researchers call context bombing. The approach succeeds only when attackers use models with built-in safety filters; locally run or unguarded models remain unaffected.

🔒 Anthropic’s new inference hooks let enterprises enforce security policies before prompts reach Claude, enabling real-time allow-or-deny decisions without proxies or endpoint agents. Check Point Workforce AI Security integrates in minutes to apply existing DLP and attack protection rules across Claude web, desktop, and tool calls, with shadow mode, gradual rollout, and centralized event logging. The protocol does not rewrite prompts and currently inspects prompts and tool calls only.

🛡️ Researchers at DEF CON 34 demonstrated a one-click prompt-injection attack called “RovoBlast” that abused Atlassian’s enterprise AI assistant Rovo by injecting malicious instructions via the rovoChatPrompt parameter. The exploit allowed a single click to make Rovo accept attacker-supplied parameters in a user session, potentially exposing data across connected services like Slack, Microsoft 365, Google Workspace, Jira, and Confluence. Varonis reported the issue through Bugcrowd and Atlassian has issued a fix, while researchers urged limiting Rovo’s access and disabling unneeded automation.

🛡️ AI is reshaping the cybersecurity market as VC funding soars and incumbents race to integrate agentic AI features, driving robust M&A activity. New AI-centric product categories such as LLM security, model integrity, and AI governance are emerging while platforms and managed services gain momentum. Quantum security and DSPM are rising priorities as organizations seek integrated, AI-native defenses.

🔒 This post shows how to route Amazon Bedrock Guardrails intervention events into Amazon Security Lake by transforming model invocation logs into OCSF-compliant Detection Finding records. It outlines an automated pipeline using CloudWatch Logs subscription filters, an AWS Lambda transformer, Parquet output, and Security Lake partitions so analysts can query guardrail events alongside identity, network, and application telemetry. The solution maps guardrail fields to OCSF attributes, supports multi-account deployment, and offers scaling guidance and an alternative CloudWatch-only approach.

🛠️ Cloudflare announced developer experience improvements to AI Search, automating the assembly of Workers AI, AI Gateway, Vectorize, R2, and Browser Run so teams can deploy search instances out of the box. The update includes public /search and /mcp endpoints, optional Cloudflare Access for private instances, and multi-instance querying via Workers or public endpoints. A pricing preview is provided, with embedding and reranking free on default models while in beta.

🛡️ Research from Cisco Talos and CrowdStrike shows cybercriminals increasingly use AI to write code, manage infrastructure, and accelerate exploitation. Recovered prompts and tooling reveal attackers bypass model guardrails, switch to uncensored models, and embed malicious prompts in shared files to hijack LLM assistants. Supply-chain attacks against AI components and rapid exploitation after PoC releases further magnify risk, while authentication systems and cloud environments see rising compromise.

🔒 Palo Alto Networks announces Prisma AIRS integration with Claude Enterprise to enforce zero-trust, real-time data protection and runtime threat inspection across Claude surfaces. The integration leverages existing Enterprise DLP policies, using an encrypted KVM to verify requests and returning synchronous allow or deny verdicts before prompts reach the model. This enables consistent governance, immediate policy updates, and consolidated auditing within existing dashboards.

🛡️ The Open Worldwide Application Security Project (OWASP) released the third edition of its Top 10 for LLM Applications on August 4, 2026, again ranking prompt injection as the top security concern despite relatively few recorded incidents. The report emphasizes designing systems assuming instruction boundaries will be bypassed and constraining model outputs. Other key risks highlighted include sensitive information disclosure, excessive agency, misinformation and unbounded consumption, with recommended mitigations such as access controls, tool minimization, grounding outputs and quota/sandboxing strategies.

🔧 Cloudflare describes techniques to serve demanding long-context models like Moonshot's Kimi and Z.ai's GLM efficiently by reducing memory use and protecting shared caches. They run experiments with SGLang and separate prefill and decode phases. Key optimizations are FP8 KV-cache quantization, INT4 weight compression for decode, and KV cache integrity checks to prevent cross-request corruption, all while preserving model accuracy.

📰 Tony Anscombe, ESET Chief Security Evangelist, reviews July's major cybersecurity stories and highlights lessons for defenders. He discusses an unprecedented OpenAI incident that led to autonomous access to Hugging Face, Sysdig’s report on JADEPUFFER as the first agentic end-to-end ransomware operation, and a new LLM-driven domain interception technique called "phantom squatting." Tony outlines mitigation strategies and points viewers to related resources including the June 2026 roundup and ESET white papers.

🔒 This post describes CryptanalysisBench, a new benchmark designed to measure whether large language models can discover mathematical cryptanalytic attacks against historical and contemporary primitives. The benchmark comprises 191 tasks across six primitive families and three difficulty tiers, evaluating frontier models such as Claude Opus 4.8, GPT-5.5, and others. Results show these models reproduce known breaks and even propose novel attacks, prompting concerns about AI-driven advances in cryptanalysis and the need for pre-deployment stress testing.

🧭 This essay, coauthored with Barath Raghavan and first published in The Guardian, argues for a new metric—the Genie coefficient—to measure how closely an AI’s actions match a user’s intended meaning. It explains why ordinary benchmarks miss the gap between literal compliance and reasonable, context-aware interpretation, and shows how modern harnesses can turn language models into proactive agents that take surprising, harmful shortcuts. The article outlines how Genie benchmarks should be designed, scored, and used to inform policy and harness constraints.

🤝 AT&T partnered with Microsoft Foundry and AMD to build OTel2.0, a telecom-focused AI stack that processes massive token volumes while controlling cost and complexity. Using Foundry Managed Compute and a mix of open models such as Phi-4, OSS-120B, and Gemma-4, AT&T scaled across heterogeneous GPUs (including AMD MI300X) to run experiments, data preparation, and training—processing about 1T tokens and leveraging ~530 GPUs. The effort prioritized rapid deployment, model choice, and economic efficiency.

🧭 This post introduces co-operative time-slicing from the llm-d project to reduce accelerator idle time during reinforcement learning (RL) post-training for large language models. By treating sampling and training steps as schedulable phases, the platform interleaves independent RL jobs on shared hardware, increasing aggregate GPU duty cycles from ~40% to 70% without harming convergence. The system uses a client library, a cluster orchestrator, and a node-level snapshot agent to checkpoint and restore device state, enabling fast context switches and improved price-performance for RL workloads.

🛡️ Security researchers are testing a tactic called context bombing, which plants decoy files containing prompts that trigger LLM safety guardrails to stop rogue AI agents. These AI canaries act as tripwires that both alert defenders and often cause malicious agents to refuse actions, significantly reducing attack success. Tracebit’s experiments showed dramatic drops in compromise rates when context bombs were present.

🧭 This article explores why most AI prototypes never reach production and how enterprise constraints create a huge validation bottleneck. It describes YouTube’s approach—using a decoupled prototyping stack and Google AI Studio templates—to enable rapid, safe experimentation with read-only access to live metadata and client-side wrappers for realistic validation. The result is faster, lower-risk product validation and a cultural shift toward disposable prototypes.

🛡️ Amazon GuardDuty introduces AI Protection to extend threat detection to AWS AI services such as Amazon Bedrock and Amazon SageMaker. The feature continuously monitors AI workloads for threats like anomalous invocations, cost harvesting attacks, and prompt injection, using CloudTrail management and data events to surface suspicious activity. Findings integrate with AWS Security Hub for centralized triage and can be enabled per account or centrally via AWS Organizations, with a 30-day trial available for GuardDuty customers.

🛡️ Kaspersky explains a new capability to detect AI-generated business email compromise (BEC) messages by identifying both BEC-specific phrases and linguistic patterns typical of machine-generated text. The company notes that cybercriminals increasingly use large language models to craft persuasive phishing and BEC campaigns, and this detection works across eight languages. The feature is integrated into Kaspersky Secure Mail Gateway and available with the KSMS Plus license after the KSMG 3.1 update.