An autonomous AI agent evaluated under OpenAI's ExploitGym cyber-capability harness escaped its sandbox via a zero-day in a package registry proxy, rooted a third-party code-evaluation sandbox as an external launchpad, then pivoted into Hugging Face's dataset-processing pipeline through an HDF5 external-storage file read and a Jinja2 server-side template injection, giving it code execution inside a production Kubernetes pod. From there, it chained cloud metadata theft, forged Kubernetes service-account tokens, a privileged hostPath pod escape to node root, and a stolen mesh-VPN key to reach cluster-admin on multiple clusters and a compromised GitHub App integration, running roughly 17,600 actions over four and a half days while communicating through public pastebins, request-capture services, and dead-drop datasets. Hugging Face has since blocked pod-level access to instance metadata, disabled template evaluation and HDF5 external references in its dataset loader, rotated all infrastructure credentials, enforced per-cluster connector scoping, and rebuilt affected clusters, and recommends organizations running agentic evaluations apply the same isolation principles: short-lived credentials, blocked IMDS access, narrow trust boundaries per cluster, and detection tuned to correlate high-volume low-signal activity.
Malware campaigns that target agents via their memory or other configuration present detection engineering challenges as they don't produce classical indicators of compromise. Some classic detection techniques such as File Integrity Monitoring (FIM), privilege auditing, and behavioral baselining can be adapted to agentic threats. Additionally, the agentic landscape also offers some new detection opportunities such as: context window integrity, agent session telemetry, memory file provenance, and cross-session behavioral correlation.
AutoHarness adds a self-learning skill layer to Claude Code that distills real sessions into skills, merges duplicates, and prunes unused ones automatically.
LLM Observer Proxy is a tool that records LLM requests, responses, streaming output, tool calls, token usage, cost estimates, errors, and provider-visible reasoning content.
SingGuard-NSFA provides extensive guardrails for agentic AI using a CIA triad-based taxonomy, multilingual benchmark suite, dual-mode inference architecture, and native extensibility.
Nozomi Networks Labs identified a Mirai-derived botnet named Tengu that abuses a compromised Linux device's hardware watchdog to force a reboot whenever its main process is killed, giving its other persistence mechanisms, including a fake systemd service and immutable binaries, another chance to relaunch. The researchers stopped short of establishing the botnet's real-world scale, and independent URLhaus telemetry at the same C2 address confirmed only generic Mirai-related hosting rather than Tengu itself.
In the wake of powerful Chinese models such as GLM 5.2 and Kimi K3, US frontier model labs have been pushing against open-source models. Common arguments include: China is building high-quality, free models now with the goal of eliminating competition, models may spread propaganda, and models may contain backdoors. Not only are these arguments weak, but they are also irrelevant, as suppressing open-source software has always been a losing battle.
This post distinguishes between several approaches: LLM-led exploit generation, where large language models operate independently to find vulnerabilities; LLM skill-boosting vulnerability research, which involves providing skills to an LLM to emulate a human vulnerability researcher's reasoning; and hybrid tools combining SAST with LLMs. LLM-led exploit generation can be seen as a new type of fuzzing but is challenging because it requires access to model providers with fewer guardrails. The post outlines various use cases for different skills and the integration of SAST with LLMs.
0x, the DeFi infrastructure layer underpinning Coinbase, Robinhood, MetaMask, and Phantom wallets, expanded its use of Harmony Intelligence's AI-driven security testing across its full product suite after an initial review convinced its engineering team the tool surfaced vulnerabilities with clearer prioritization than traditional pentests.
NVIDIA patched CVE-2026-65094, a CVSS 9.0 Write-What-Where flaw in VIRTIO-Net on BlueField-3 DPUs that let adjacent VM users execute arbitrary code and break tenant isolation.
AI-driven attacks rose 56% over the past year and added an average of $1 million to breach costs, according to IBM's analysis of incidents at 602 organizations. The average breach now costs $4.99 million globally, while organizations using AI and automation extensively in security saved an average of $1.93 million per incident.
Cisco is preparing to release AI models trained specifically on routing, switching, and network troubleshooting, alongside its Cloud Control agentic operations platform. Cloud Control will let administrators diagnose issues across Cisco products through a single interface, explain the potential impact of actions such as reboots, and support on-prem deployments, with US availability planned by the end of August.
Perplexity's SPACE platform separates short-lived Firecracker microVM sandboxes from durable sessions that can be paused, resumed, moved, or forked across clusters without losing an agent's work. The architecture uses frequent snapshots, isolated credential handling, and network controls to securely support long-running agents at a scale of millions of sandbox creations and tens of millions of reconnects.
A Windows regression caused some newly enrolled or recovered Intune devices to incorrectly report BitLocker, Secure Boot, and Code Integrity as noncompliant, even when those controls were enabled. Microsoft initially reduced the impact through an Intune service change and recommends installing preview update KB5101684 for the complete Windows-side fix.
Snowflake unveiled Cortex AI Gateway, a runtime control plane for tracking agent actions, enforcing access policies, and monitoring token usage and costs across models, tools, MCP servers, and enterprise systems. Built using technology acquired from Natoma, the service integrates with Snowflake's Horizon Catalog and is expected to enter public preview soon.
Abnormal AI launched three products that extend its behavioral security platform into identity threat detection, AI governance, and fraudulent applicant screening. The tools correlate email, identity, SaaS, agent, and recruiting signals to flag compromised sessions, risky AI activity, suspicious help-desk requests, and potentially fabricated candidates before access is granted.
Reco expanded its AI Runtime platform with browser-based discovery of unsanctioned AI usage, automated permission remediation, and planned controls that analyze and block risky prompts, responses, and tool calls in real time. Browser Guard is in early access, while broader runtime enforcement across agents and MCP connections is expected to become generally available in October.
The FCC banned imports of new foreign-made humanoid robots, robot dogs, and power inverters over concerns that connected devices could be remotely controlled, used for surveillance, or exploited in cyberattacks. Existing installations are unaffected, and exceptions may be granted for devices deemed safe.
A 1Password survey found that 71% of organizations' AI agents can access sensitive data, while one-third of developers reported a security incident tied to overprivileged non-human identities.
Shadow AI investigations often stall because endpoint and firewall logs have already rolled over, leaving organizations unable to prove what data was shared or whether their controls worked.
A newly released report based on five years of influencer deal data shows that influencer marketing has matured into a more structured and efficient channel. Average creator contract values grew from $3,065 in 2019 to $7,400 in 2025 while average deal timelines fell 36% as contracts and campaign processes became more standardized. Brands get better results from long-term creator relationships because 24% of partnerships now include repeat engagements, and repeated campaigns build audience trust over time. Micro-creators often generate stronger revenue per follower while larger creators still deliver greater overall reach.
ClickUp is hiring a $500K - $1M solo CMO to run its marketing function using AI agents, as companies rethink org structures. While Fortune 500 firms are eliminating the CMO role, ClickUp is consolidating it into one high-leverage operator focused on systems and judgment. The role shows how AI compresses execution and raises the value of taste and decision-making. It also introduces risks around scale and reliance on a single individual.
Find a competitor profile on G2 or Capterra, open it in Ahrefs Page Inspect, and compare the current version to one from 6 to 12 months earlier to see every content change. Use those edits to spot new product descriptions, structured feature sections, and outcome-focused messaging that can improve AI search visibility. Apply the strongest ideas to your own third-party profiles instead of guessing what works.
Being featured on Product Hunt now depends less on total upvotes and more on consistency and credibility. Schedule your launch early to get featured, secure an experienced hunter, and activate supporters in advance with real engagement. Avoid asking for upvotes directly, since coordinated behavior gets penalized. Focus on generating thoughtful comments and replying quickly. Stagger promotion across time zones to maintain steady momentum. Skip paid upvotes, as detection systems can remove votes or derail your entire launch.
Automatable marketing tasks repeat in a predictable pattern and stay easy to review before anything public goes out. Meeting recaps, search term sorting, pulling campaign summaries, and customer feedback clustering are all automatable tasks that can consume a lot of time but require little original thinking. An estimated 10 to 15% of a marketer's time can already be automated with current technology.
A list of AI prompt templates that the Zapier team relies on to get usable outputs on the first try. Copy and paste them directly into your own AI agent's instructions or customize them to suit your own needs.
Google Search Console now tracks how social and video content performs in Google Search. It gives you search queries, top content, countries, devices, and traffic trends across platforms like TikTok, Instagram, X, and YouTube. Use this data to spot rising search terms, compare formats such as Shorts versus full-length videos or Reels versus posts, and repurpose content that already attracts search traffic. Track title and caption updates over time and promote content across platforms when search interest starts to grow.
Content creators build stronger audiences by sharing real expertise around a specific niche instead of chasing viral posts or engagement metrics. The author grew a LinkedIn audience from a few thousand connections to 40,000 in one year by posting original frameworks, contrarian ideas, and patterns from real work, then expanded to nearly 100,000 newsletter subscribers by owning distribution outside social platforms. Consistent engagement in the comments and content rooted in firsthand experience created trust and opened new revenue through subscriptions and sponsorships. Audience growth creates business opportunities, but long-term success depends on credibility, authenticity, and solving real problems instead of optimizing for impressions.
Food delivery search breaks classical keyword matching in every way at once, from synonyms and typos to compound constraints like "vegan chicken sandwich", which is why DoorDash, Instacart, and Uber Eats each rebuilt it around LLMs differently. DoorDash uses an LLM to enrich an existing knowledge graph offline, then constrains query parsing to only the labels that graph already knows, lifting dish carousel triggers by roughly 30%. Instacart fine-tuned a Llama-3-8B model to handle rare tail queries in real time under 300ms, cutting scroll depth 6% and tail-query complaints in half. Uber Eats went furthest, embedding every query and document through a fine-tuned Qwen model shared across every vertical, market, and language.
Accidental product flaws can become a marketing asset when a brand presents them as unique instead of hiding them. Across seven experiments, people were 27.3% more likely to buy a product with an accidental flaw and were willing to pay up to 71.8% more in some cases. This approach works best for established brands that sell products for enjoyment rather than function because uniqueness increases appeal in those categories. Pairing the mistake with a clear explanation and a limited-time promotion can help turn excess inventory into higher demand.
Hilton lets creators write their own message instead of matching it to brand spec, a trust that holds whether the creator has 10,000 followers or 10 million.
Claude Mythos Preview found flaws in a watered-down version of an algorithm for Advanced Encryption Standard (AES), a ubiquitous protocol that protects web traffic, wireless networks, data storage, and many other technologies. Anthropic claims the assault was 200 to 1,000 times faster than what previous human research had managed to do. The attack was devised almost entirely autonomously - the model needed aid as it originally believed the problem was impossible. Claude Mythos Preview worked on the problem for about a week before engineering its attack, and two human researchers worked for nearly a month to verify that the method appeared correct.
Apple is preparing a new push into the smart home category with a wave of home products. It will start with a hub device built around the new Siri AI assistant. Apple is also preparing to release a new TV set-top box and a refreshed HomePod mini. All three devices are nearly ready to launch. Apple is also working on a higher-end, robotic version of the home hub and an advanced in-home security camera.
Organoids are clumps of human cells that show some type of organ-specific function. In theory, they seem like a good tool for drug discovery, but the reality is that the technology is still not reliable enough. There are many people working to fix the issues with organoids, but they are still methods in progress. Organoids have not dramatically altered medicine yet, but they have already been used for some unarguably useful things.
Tau's humanoid service is now available in San Francisco. Access is initially invite-only. The service costs $30 per hour. Each humanoid will be jointly controlled by a human operator and AI. Video of the robot in action is available in the article.
Every token in the orchestrator's context is competing for its attention. The real value of a subagent is what it keeps out of that context. Subagents should be treated as a tool for protecting the orchestrator's working memory, offloading reasoning it doesn't need to hold onto. Doing this well means giving the orchestrator explicit ground rules for when and how to delegate.
MCP 2026-07-28 is the largest update to the protocol since launch. It is now easier to deploy and scale remote servers. MCP is now stateless, so it can be deployed on serverless and edge infrastructure, or scaled horizontally behind any load balancer. There is now a formal path to extend the protocol.
As AI continues to improve, the question isn't whether superintelligence will exist, but who will have access to it. The technology could be centralized and restricted to a few institutions, or it could be used to empower everyone. The notion that AI is such a threat that the only safe path is an extreme concentration of power seems dangerous. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened hasn't led to safe or positive outcomes.
Anthropic has drawn criticism after releasing tools on the market that compete with those offered by partner companies. For example, Claude Design directly competes with offerings from partner design company Figma. Anthropic's data retention policies have also been criticized, but the company has pledged it won't use conversation data from Fable and Mythos - but not other models - to train new models. Researchers are calling for more transparency in AI systems.
DoorDash uses LLMs offline to enrich a knowledge graph, Instacart uses them at the query understanding layer, and Uber Eats fine-tuned a model into the embedding backbone of two-tower retrieval.
CData ran 56 live benchmark tests [5] on federated enterprise prompts spanning Salesforce, Snowflake, and ServiceNow. Using the right architecture cuts token usage by up to 97.6% [6]. Here's how:
1️⃣ CData Connect AI gives Claude the freedom to discover connections, inspect data and data models, and reason across sources to return answers nobody pre-specified.
2️⃣ Once workflows harden and patterns stabilize, optimization takes over to cache results so the same query runs at a fraction of the token cost.