Settings

The log

The time column takes an ISO timestamp, a date and time separated by a space, or a bare epoch in seconds or milliseconds. A connection log is a map of an entire organisation; this one is read by the browser on this device and is not sent anywhere.

What has the shape

Period is the median gap between connections. Jitter is how much the gaps vary, doubled, because an implant sleeping for a set time plus or minus a percentage produces gaps spread evenly over that range, whose median absolute deviation is exactly half the percentage. In band is the share of gaps landing close to the median. Size cv is how much the conversation size varies, where zero means every check-in is identical.

The gaps, drawn

The analysis


    
How it works

Malware that calls home to a command server does it on a timer, and nothing a person does is regular to the second. BEACON reads network connection logs and finds hosts contacting the same destination at steady intervals, allowing for jitter, using a robust measure of spread rather than an average that one outlier can skew. It flags candidates for a person to check; it does not decide what is malicious. The logs stay in this tab.

The guides: How to spot malware beaconing in a connection log and How to find generated domains in a DNS log

Questions people ask about BEACON

Why does jitter not defeat this?

Because jitter has a shape. An implant told to sleep sixty seconds with twenty per cent jitter produces gaps spread evenly over forty-eight to seventy-two seconds, and a spread like that has a median absolute deviation of exactly a quarter of its width. Divide that by the median and you get half the jitter percentage back, every time. So rather than being noise that hides the beacon, the jitter is a second measurement of it: the page reports the sleep and the jitter separately, which is what goes in the report and what gets matched against another incident.

Why the median absolute deviation and not a standard deviation?

Because of the laptop that gets closed at lunchtime. A beacon that checks in every five minutes and then misses an hour produces one gap of 3,600 seconds among hundreds of 300. That single value moves a standard deviation enough to hide the beacon completely, and moves a median absolute deviation by nothing at all. The page counts those long gaps separately and calls them missed check-ins, because how often a machine is switched off is itself worth knowing.

Will this find the beacon or tell me it is malicious?

The first, never the second. A backup agent, a monitoring check, a software updater and a certificate refresh are all beacons by every measure on this page, and in most networks they will fill the top of the table. Nothing here can tell them from an implant, because nothing in the timing distinguishes them: the difference is what the destination is, and that is your judgement. Any tool that prints a verdict here is selling you the one thing it cannot do.

What does the score actually mean?

It is half regularity, just over a third whether there is only one interval, and a sixth whether the payload is constant. Every one of those three is a number between nought and one that is printed in the same table, so the score can be recomputed by hand from the columns next to it. That is deliberate: a score nobody can check is a score nobody should act on, and there is a great deal of security software that has never been asked to show its working.

What format does the log have to be in?

Comma or tab separated, with a time, a source and a destination. That covers a Zeek conn.log, a firewall or proxy export, a NetFlow dump and most SIEM exports; you tell it which column is which rather than it guessing. A byte count makes the payload measure work and a port makes the results easier to read, but neither is required.

How long a window should I give it?

Long enough to hold at least a few dozen check-ins of whatever period you are looking for, which for an hourly beacon means several days rather than one. Too short and there are not enough gaps to measure; too long and a machine that changed its behaviour halfway through reads as irregular when it was two regular things in sequence. If a pair scores oddly, narrowing the window and running it again usually explains why.

Doing this often? The Sovereign Console loads logs, captures and files into one case, pulls every indicator and timestamp, and exports a report. On its own, US$2 a month; with everything else, US$5 a month or US$48 a year.