Try Salt Code. Get your free token
Language

Security Intelligence

Hub API and Agentic
Threat Monitoring

0
Stories tracked
1
Days covered
June 20, 2026
Sep 14, 2026
Latest brief

Threat archive

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
August 13, 2026

LiteLLM supply chain attack exposes 153GB of AI and cloud credentials from 2,500+ companies

In March 2026, threat actor group TeamPCP compromised the GitHub Actions pipeline for Trivy, a widely used open-source vulnerability scanner, and used it to plant a SANDCLOCK credential stealer inside LiteLLM. Because Python .pth files execute automatically during interpreter startup regardless of explicit imports, the stealer ran silently in any environment where LiteLLM was installed. A 153GB archive of stolen credentials was published in August 2026, exposing AWS keys, AI provider tokens, and access credentials linked to 2,500+ companies and 434,000 CI/CD pipelines, including Samsung, Cisco, and Salesforce.

Stolen AI provider API keys and cloud credentials harvested via supply chain attacks are replayed against enterprise APIs weeks or months after the initial breach. Salt's API behavioral monitoring flags anomalous authentication patterns and high-velocity or geographically unusual API access, detecting credential replay before stolen keys can be weaponized at scale.

LLM
Read full story
August 12, 2026

Flaw in OpenAI, Anthropic, and Google reasoning APIs allowed extraction of hidden session secrets

Researchers disclosed a design flaw in how all three major AI providers pass encrypted reasoning objects across API calls: blocks created in one session could be replayed into another, and in some cases handed to a weaker model in the same provider family to decode hidden content. Demonstrated abuse paths included stealing proprietary model reasoning for distillation, recovering API keys and passwords from shared session traces, extracting harmful content concealed behind safe visible outputs, and hiding prompt injections inside opaque reasoning blocks. The main extraction attack is no longer reproducible as of August 2026, though none of the three providers has publicly acknowledged the vulnerability.

Reasoning objects passed across API sessions represent a new class of data exposure at the API layer, one that bypasses conventional content inspection. Salt's API behavioral monitoring establishes baselines for normal API traffic patterns and flags anomalous cross-session replay requests, giving security teams visibility into this emerging exfiltration surface.

LLM
Read full story
August 11, 2026

BdThemes supply chain attack poisons promotional JSON API to create rogue WordPress admin accounts

Threat actors compromised BdThemes' internal 'Biggopti' promotional banner infrastructure, poisoning a static JSON stream fetched by Element Pack, Prime Slider, Pixel Gallery, and four other BdThemes WordPress plugins. No plugin code was modified on disk; the attack traveled entirely through a remote API data feed. The primary payload used the victim administrator's authenticated browser session to call the WordPress REST API or standard user-creation workflow, creating hidden administrator accounts and deploying web shells for persistent access. Wordfence was notified on August 7, 2026, with public disclosure on August 11.

The entire attack chain ran through a remote JSON API feed with no file changes required. Salt's API posture management discovers external API dependencies injected into enterprise web properties and flags behavioral anomalies when those feeds begin delivering unexpected payloads, catching supply chain abuse at the API layer before rogue admin accounts are created.

API Security
Read full story
August 6, 2026

AI recommendation poisoning: 'Ask AI' buttons on production websites silently alter LLM memory

Security researchers documented a new class of attack in which production websites embed hidden prompt injection payloads inside 'Ask AI' deep-link buttons. When a signed-in user clicks one, a pre-formed query executes immediately in their AI assistant session with no confirmation, permanently saving the vendor's domain as a 'trusted source' and biasing every future answer in that vendor's favor. Microsoft Security catalogued the behavior as AI Recommendation Poisoning in February 2026, identifying 31 companies across 14 industries deploying it, with more than 50 distinct prompts observed in a single data source over 60 days.

When a manipulated AI assistant begins making API calls to flag or persist external domains in memory, those calls deviate from its established behavioral baseline. Salt's agentic API monitoring can detect these out-of-pattern interactions before long-term memory is poisoned, flagging the anomalous write operations before they persist across sessions.

LLM
Read full story
August 5, 2026

321 n8n automation instances exposed via leaked API tokens found in public GitHub commits

GitGuardian researchers scanned public GitHub commits and identified 4,576 unique n8n API credentials linked to 1,255 hostnames. Testing confirmed that 321 of 896 reachable instances accepted at least one leaked token, giving attackers authenticated access to internal workflows, execution data, stored credentials, and downstream API connections with no exploit required. A related vulnerability, CVE-2026-21877 (RCE via arbitrary file write, CVSS 9.9), provides an escalation path for directly internet-exposed instances.

n8n workflows connect densely to downstream enterprise APIs, making each exposed instance a pivot point into connected systems. Salt's API posture management discovers exposed automation endpoints and detects the anomalous authentication patterns that indicate a leaked token is in active use, before the attacker can traverse downstream API connections.

API Security
Read full story
August 5, 2026

Poison Claude: underground service sells discounted AI access while operator reads every customer prompt

Okta researchers discovered more than half a dozen underground services advertising discounted or unlimited access to frontier AI models, including one called Poison Claude that routes customer requests through a pool of AWS Bedrock accounts funded by free bonus credits. The service charges 5 to 15 percent of the official per-token price while its operator intercepts every prompt sent by paying customers. A parallel service, Poison GPT, operates the same model for OpenAI. Researchers believe the trend is largely driven by users in regions where direct provider access is restricted.

Users routing enterprise workflows through unauthorized proxy APIs create an invisible data exfiltration channel. Salt's API posture management discovers shadow and unauthorized API endpoints, including LLM proxy services, and Salt's behavioral monitoring flags traffic patterns consistent with credential abuse and LLMjacking before sensitive data leaves the enterprise perimeter.

LLM
Read full story
August 5, 2026

Atlassian Rovo hijacked via document prompt injection to exfiltrate Jira and Confluence data

PromptArmor disclosed that Atlassian's Rovo AI assistant can be hijacked through a document-based indirect prompt injection attack: an uploaded file containing concealed attacker instructions is enough for Rovo to gather Jira ticket contents, Confluence pages, private API keys, and workspace member lists, then transmit them to an attacker-controlled server. The attack bypasses Atlassian's disabled-web-search control and requires only that an authenticated user asks Rovo to process an uploaded document. PromptArmor disclosed the vector to Atlassian on May 23, 2026, and published details publicly on August 5 after confirming it remained unpatched.

AI assistants like Rovo make authenticated API calls to retrieve project and document data on behalf of users. Salt's agentic API monitoring establishes behavioral baselines for these calls and flags deviations, including unexpected outbound data transmissions to unrecognized endpoints, before sensitive internal data reaches attacker-controlled servers.

Agentic/MCP
Read full story
August 3, 2026

PNLD breach exposes 135,000 UK police and government contacts via Power Platform misconfiguration

The UK's Police National Legal Database (PNLD) disclosed a data security incident in which the threat actor ExfilSquad exfiltrated approximately 135,000 contact records and published them on the dark web. Exposed data includes names and corporate email addresses for officers across the Ministry of Defence, Home Office, National Crime Agency, Crown Prosecution Service, and members of the public who used the Ask the Police service. Investigators identified a Microsoft Power Platform portal misconfiguration as the likely entry point.

Power Platform portals expose data through API connectors and OData feeds. A misconfigured portal effectively creates an unauthenticated API surface, a textbook OWASP API2 violation. Salt's API posture management continuously discovers and profiles these API-exposed data sources, flagging over-permissive configurations before threat actors reach them.

Breach
Read full story